"The suit seeks nothing less than the erasure of both any GPT instances that the parties have trained using material from the Times, as well as the destruction of the datasets that were used for the training. It also asks for a permanent injunction to prevent similar conduct in the future. The Times also wants money, lots and lots of money: "statutory damages, compensatory damages, restitution, disgorgement, and any other relief that may be permitted by law or equity.""
If we see court judgements start to go copyright owners way, we will also see a scramble from AI companies to buy the few publishers with enough data to be worth buying, and to create works for hire to replace the rest.
In the long run a copyright ruling like that will be a boon for OpenAI and all other players with deep enough pockets to do so, and massively harm everyone else who will suddenly find it far harder to build models legally.
So that would mean articles from the 1920s, provided that the authors of those articles have been dead for 70 years, or longer in some other countries.
On the other hand, NYT website willingly gave out all the information without imposing limitations. Seeing terms of service requires visiting a separate page, they aren't seen immediately upon visiting the website. Understanding and accepting the terms also requires a human interaction.
robots.txt on nytimes.com now disallows indexing by GPTBot, so there's an argument against automated information acquisition starting from some moment, but before some moment they weren't explicitly against that.
> Seems weird to argue that you have to speak up if you don’t want something done to you or else you consent to everything.
If you don't want people to get at your land, setting up even a small fence creates an explicit indication of limitations. Just like the record in robots.txt I mentioned earlier.
New York Times also doesn't limit article text content if you just request HTML, which is typical for automated cases. But they impose th limits imposed on users viewing the pages in browser with Javascript, CSS and everything else. So they clearly:
1. Have a way to determine the user's eligibility for reading the full article on server side.
2. Don't limit the content for typical automated cases on server side.
3. Have a way to track the activity of not logged in users, determining the eligibility for access. So it's reasonable to assume that they had records of repeated access from the same origin, but didn't impose any limitations before some time.
So there are enough reasons to think that robots are welcome to read the articles fully. I'm not talking about copyright violations here, only about the ability to receive the data.
It's obviously a frivolous suit that will only net at best a ceremonial victory for NYTimes: 8 figure max payout and a promise to not use NYtimes material in the future.
The trajectory and value to society of OpenAI vs NYtimes could not be greater. They have won no favors in the court of public opinion with their frequent misinformation. It's all just a big waste of time, the last of the old guard flailing against the march of progress.
And even hypothetially if they managed to get OpenAI to delete ChatGPT they'd be hated forever.
ChatGPT only advertises itself as a fancy autocomplete. There is a disclaimer that it may produce output that appears correct but isn't. NYtimes written material purports itself to be the truth, thus shouldn't be held to the same standards as a generative AI obviously.
I think what we should focus on is the volume of misinformation in general, not the provenance of it.
The NYT may produce misinformation but it aims not to, and its staff of human writers are limited in the quantity that they can produce. They also publish corrections.
GPT enables anyone who can pay to generate a virtually unlimited volume of misinformation, launder it into 'articles' with fake bylines and saturate the internet with garbage.
Well that's true for any large language model. As long as they exist there will be a deluge of bot written text producible for any purpose. At this point there is no getting the cat back into the bag.
In that case the bigger danger is Open source LLM's. OpenAI at least monitors the use of their endpoints for obvious harm.
> The NYT may produce misinformation but it aims not to, and its staff of human writers are limited in the quantity that they can produce. They also publish corrections.
Except when it affects their bottom line of course, they publicly lied on how meta tags work during the lawsuits against Google to get more money (like most newspapers did). And I have no doubt that they will extensively lie once again on how LLM really work.
Something I have wondered about LLMs and training data is the idea that the biggest content producers on the internet now have their world view and tone echoed disproportionately as part of the next big wave of technology. This is incredibly impactful (although admittedly I don't know how to turn that into a profit). Is there some long term impact of removing the New York Times from training data that means it won't be part of the LLMs corpus going forward that is unforeseen?
So presumably when they fix that issue (which, if the text matches exactly, should be trivially easy) then would you accept that as a sufficient remedy?
Copyright infringement is not avoided by changing some text so it isn’t an exact clone of the source.
Determining whether a work violates a copyright requires holistic consideration of the similarity of the work to the copyrighted material, the purpose of the work, and the work’s impact on the copyright holder.
There is not an algorithm for this, cases are decided on by people.
There are algorithms that could detect obvious violations of copyright, such as the one you suggest which looks for exact matches to copyrighted material. However, there are many potential outputs, or patterns of output, which would be copyright violation and would not be caught by this trivial test.
I certainly don't think it's impossible, but I think it is hard problem that won't be solved in the immediate future, and creators of data used for training are right to seek to stop wide availability of LLMs that regurgitate information they worked hard to obtain.
I think it will be a bit easier than you believe. The reason why it hasn’t been done yet is that there hasn’t been a compelling economic reason to do so.
In Sony vs. Universal case, Sony is the producer of a tool where the consumer uses to "time-shift" a broadcast that they legally are allowed to view. Similarly, you can rip your own CDs or photocopy your own books. This case never made reselling those content legal. OpenAI does not train ChatGPT on the content you own - they do it on some undisclosed amount of data that you may or may not have a legal right to access, and then move on and (is shown to) reproduce it nearly verbatim - they may even charge you for the pleasure.
NYT won't mind if you use their content to train LLMs - as long as they get a commission. Reddit will shut down their free API and make you pay to get training content. Discord is going to be selling content for AI training too - if they haven't already done so. Twitter is doing it.
They didn't care before because LLMs were just experiments. Now we're talking trillions of dollars of value.
> They didn't care before because LLMs were just experiments. Now we're talking trillions of dollars of value.
Can you make the argument this was their fault for not having forward vision/being asleep at the wheel and "accidentally, in hindsight" letting OpenAI/others have free, open, unlimited access to their content?
Basically none of the training material for GPT was used under an "unlimited" license. There are very important legal limitations. GPT just doesn't care much about them.
"They" also include the people working there. Why someone work with full time writing articles should give the work for free just let someone to train it and make money out of it as a consequence?
>Why someone work with full time writing articles should give the work for free
They are not giving it out "for free", in fact they're being paid by their employer to write these articles. Moreover, the writers themselves stand noth' to gain from their past writings financially as they don't belong to the ownership structure of the business.
Their ability to make money in the future is directly tied to their employers' ability to make money with their content. This is a closed financial loop. If OpenAI or any other AI company wants in, they should pay a licensing fee or get the laws changed, not just assume that they can take what they want and pretend like there are no negative consequences for the creator or the rights-holder.
In this limited example, are there such consequences? Are people dropping NYT subscriptions because they trust chatgpt to inform them of current events? I don’t buy it.
No one is pretending there are no "there are no negative consequences for the creator or the rights-holder". Of course there are. But this is a story of rights-holders, who've already outgrown their usefulness, wanting to tap themselves into money stream they are not entitled to.
ChatGPT isn't competing with NYT on a core competency. No one uses LLMs for original news reporting. They're obviously incapable of doing that, by virtue of not being there on the scene or able to independently research a topic, maintain relationships with sources, etc. What ChatGPT can do is quote/reproduce some parts of past articles, and reason from them. Or at least produce new text that's somewhat related to the old text.
The threat to NYT is this: ChatGPT is much better bullshitter than they are, so it reduces NYT to its core competency: providing original information. Which is all it should be doing in the first place. But instead, NYT wants to not only keep the bullshitting part of its revenue, but also take a cut or destroy the much greater and much more useful part of where this all feeds a general-purpose language model.
This is a badly-formulated conjecture, or worse, ultimately selective reading of "social credit" which only purpose is serving your argument; it has nothing to do with economics. I'm sorry, but I'm not convinced.
When open source developers do that, they also include an explicit licensing information that lists cases when the usage is allowed and restricted. So even if the code is open source and licensed under GPL, its usage in a closed source product like ChatGPT is not allowed.
GPL code usage in closed source ChatGPT is allowed "for internal use"; it just would not be allowed to distribute binaries of ChatGPT that are closed source without making source available; also a GPL3 license violation to allow online access to a ChatGPT program that used GPL3 code without making source available.
So Just because some people give out something for free at certain time, all the other people should do the same all the time? Not to mention most open source comes with a well-defined term not just exploited from free by a closed service making money for another company.
With the ways NYTimes has degraded since 2010 even if people there are working for free, they're still being overpaid. The only adequate section there is the food.
The journalists created the content for the NYT, the users created it for Facebook. Both received something in return for their effort, and the content ended up being owned by NYT/facebook
The journalists were paid to assign ownership of their work to the NYT corporation, with a clear and well understood contract of work that they signed, either in real ink or with an equivalent electronic signature, as consenting adults.
Can you say the same for user created content on Reddit, Twitter, or Facebook? A user agreement that nobody reads doesn't have anything like the same legal basis as a signed contract. Not to mention that a large percentage of users are not adults.
I've been arguing since ChatGPT came out that LLMs should fall under fair use as a "transformative work". I'm not a lawyer and this is just my non-expert opinion, but it will be interesting to see what the legal system has to say about this.
This tired 'fair use' excuses from AI bros whilst the GPT has reproduced the article text verbatim, word for word and it being monetized without the permission from the copyright holder and source (NYT) is an obvious copyright violation 101. Full stop.
Again, just like Getty v. Stability, this copyright lawsuit will end in a licensing deal. Apple played it smart with their choice with licensing deals to train their GPT [0]. But this time, OpenAI knew they could get a license to train on NYT articles but chose not to.
the purpose and character of the use
the nature of the copyrighted work
the amount and substantiality of the portion taken
the effect of the use upon the potential market.
Literally every single one of these factors has very complicated precedent and each one is an open question when it comes to AI. Since fair use is a balancing test this could go any way.
Stability took the easy way out because they didn't have billions of dollars to play around with and Microsoft to back them. Let's see what OpenAI does but calling everyone who disagrees with your naive interpretation of fair use "AI bros" is doing everyone a disservice.
Young males that wear Tensorflow branded muscle tank tops and drive Mitsubishi Eclipse convertibles with the vanity plate OVERFIT. They are everywhere these days.
Thank you for the absurd visual. The vanity plate, especially, was worth saving for last. Somehow, the car is well suited, also. Love how they prefer Tensorflow over Pytorch, too.
I generally tend to downvote comments that use "x bros" for pretty much any x on sight for that reason. It's exceedingly rare for such a comment to be much more than a thinly veiled insult with little substance. Sometimes I might even agree with the insult, but it's still rarely appropriate here.
I don’t doubt it does. It’s easy to get it to spit out long answers from Stack Overflow verbatim, I’ve done it. Maybe some of the “transformative” nature of the LLM output is the removal of any authorship, copyright, license, and edit history information. ;) The point here is to supplant Google as the portal of information, right? It doesn’t have new information, but it’s pretty good at remixing the words from multiple sources, when it has multiple sources. One possible reason for their legal woes wrt copyright is that it’s also great at memorizing things that only have one source. My college Markov-chain text predictor would do the same thing and easily get stuck in local regions if it couldn’t match something else.
I'm sure the NYT uses dictionaries, encyclopaedias and style books verbatim as well. And they don't invent the facts they write about. As journalists they are compiling and passing along other knowledge. You usually don't get a piece of their income when a journalist quotes you verbatim (people usually don't get paid for interviews).
It's inevitable that this question ends up at the supreme court. And the sooner the better IMO. It's clearly fair use. Generative agents will be seen legally as no different than a human artist leveraging the summation of their influences to create a new work.
It's not stored in ChatGPT actually, unlike Google's web search cache where it is stored verbatim, can be recalled perfectly, and is still fair use.
Fair use has nothing to do with reproducibility. LLMs are more clearly fair use than a search engine cache and those court cases are long settled. There's no world in which OpenAI doesn't win this entire thing.
>It's inevitable that this question ends up at the supreme court. And the sooner the better IMO. It's clearly fair use. Generative agents will be seen legally as no different than a human artist leveraging the summation of their influences to create a new work.
Why do you think the architecture is important?
If I have a computer program and it outputs the an entire copyrighted poem then the answer to "is this copyright violation" SHOULD NOT depends on the architecture of the program.
This seems like a reasonable opinion when you think about the training data size and imagine that any given output is some kind of interpolation of some unknown large number of training examples all from different people. If it’s borrowing snippets from tens or hundreds or thousands of sources, then who’s copyrights are being violated? Remixing in music seems to be withstanding some amount of legal scrutiny, as long as the remix is borrowing from multiple sources and the music is clearly different and original.
It gets harder to stand behind a blanket claim that LLMs or any AI we’ve got falls under fair use when they keep repeatedly reproducing complete and identifiable individual works and clearly violating copyright laws in specific instances. The models might be remixing and/or transformative most of the time, but we have proof that they don’t do that every time nor all the time… yet. Maybe the lawsuits will be the impetus we need to fix the AIs so they don’t reproduce specific works, and thus make the fair use claim solid and actually defensible?
36": 'however, the press as you know it has ceased to exist'
40": '20th-century news organizations are an afterthought; a lonely remnant of a not-too-distant past'
2'11": 'also in 2002, google launches google news, a news portal. news organizations cry foul. google news is edited entirely by computers'
5'13": 'the news wars of 2010 are notable for the fact that no actual news organizations take part. googlezon finally checkmates microsoft with a feature the software giant cannot match: using a new algorithm, googlezon's computers construct new stories, dynamically stripping sentences and facts from all content sources, and recombining them. the computer writes a new story for every user'
5'55": 'in 2011 the slumbering fourth estate awakes to make its first and final stand. the new york times company sues googlezon, claiming that the company's fact-stripping robots are a violation of copyright law. the case goes all the way to the supreme court'
they didn't get the details exactly right, but overall the accuracy is astounding
however, that may be a hyperstition artifact in this timeline
https://en.wikipedia.org/wiki/EPIC_2014 (i thought epic 2014 might be the only flash video to hae a wikipedia article about it, but then i looked and found five others)
The suit demonstrates instances where ChatGTP / Bing Copilot copy from the NYT verbatim. I think it is hard to argue that such copying constitutes "fair use".
However, OAI/MS should be able to fix this within the current paradigm: Just learn to recognize and punish plagiarism via RLHF.
However, the suit goes far beyond claiming that such copying violates their copyright: "Unauthorized copying of Times Works without payment to train LLMs is a substitutive use that is not justified by any transformative purpose."
This is a strong claim that just downloading articles into training data is what violates the copyright. That GTP outputs verbatim copies is a red herring. Hopefully the judge(s) will notice and direct focus on the interesting, high-stakes, and murky legal issues raised when we ask: What about a model can (or can't) be "transformative"?
Transformations are happening. Maybe if the output is verbatim afterwards, than that says something about the outputs originality all along... or am I a troll?
They're talking about transformative with regard to copyright law where it is an important part of determining fair use, not the dictionary definition you're using here.
I can't take NY Times articles, translate them into Spanish, and then sell the translations under fair use, even though clearly I've transformed the original article content.
Well yeah, copying a work and using it for its original expressive purpose isn’t fair use, no? You have to use it for a transformative purpose.
Suppose I’m selling subscriptions to the New Jersey Times, a site which simply downloads New York Times articles and passes them through an autoencoder with some random noise. It serves the exact same purpose as the New York Times website, except I make the money. Is that fair use?
> Well yeah, copying a work and using it for its original expressive purpose isn’t fair use, no? You have to use it for a transformative purpose.
They transformed the weights.
Just like reading the article transforms yours.
As for verbatim reproduction, I'm pretty sure brains are capable of reproducing song lyrics, musical melodies, common symbols ("cool S"), and lots of other things verbatim too.
Those quotes from Dr. King's speech that you remember are copyrighted, you know?
This comment is just blatant anthropomorphizing of ML models. You have no idea if reading an article “transforms weights” in a human mind, and regardless, they aren’t legally the same thing anyway.
Why? A human being isn’t infinitely scalable; they’re just different. It’s the same thing as going to a movie theatre to watch a movie vs. recording it with a camera.
A human churning butter, spinning cotton, or acting as a bank teller isn't infinitely scalable either. This is orthogonal to the point.
Times change. We're industrializing information creation and consumption (the latter is mostly here already), and we can't be stuck in the old copyright regime. It'll be useless in very short order.
All this road bump will do will give the giant megacorps time to ink deals, solidify their lead, and trounce open source. Twenty years on, the pace of content creation will be as rapid as thought itself and we'll kick ourselves for cementing their lead.
This is a transitional period between two wildly different worlds.
If they could find a single person who in natural use (e.g. not as they were trying to gather data for this lawsuit) has ever actually used ChatGPT as a direct substitution for a NYT subscription, I'd support this lawsuit.
But nobody would do that, because ChatGPT is a really shitty way to read NYT articles (it's stale, it can't reliably reproduce them, etc.). All that is valuable about it is the way that it transforms and operates on that data in conjunction with all the other data that it has.
The real world use of ChatGPT is very transformative, even if you can trick it into behaving in ways that are not. If the courts act intelligently they should at least weigh that as part of their decision.
It’s more of a thought experiment. Here’s another with more commercial applications:
Suppose I start a service called “EastlawAI” by downloading the Westlaw database and hiring a team of comedians to write very funny lawyer jokes.
I take Westlaw cases and lawyer jokes and feed them to my autoencoder. I also learn a mapping from user queries to decoder inputs.
I sell an API and advertise it to startups as capable of answering any legal question in a funny way. Another company comes along with an API to make the output less funny.
Have I created a competitor to Westlaw by copying Westlaw’s works for their original expressive purpose and exposing it as an intermediary? Or have I simply trained the world’s most informative lawyer joke generator that some of my customers happen to use for legal analysis by layering other tools atop my output?
Did I need to download Westlaw cases to make my lawyer joke generator? Are the jokes a fair-use smokescreen for repackaging commercially valuable copyrighted data? Does my joke generator impact Westlaw in the market? Depends, right?
That’s nonsense piracy. I never intend to own a truck, so when I need to haul a little something I go to Home Depot and steal a Ford off the lot for an hour? What if I stole all your commits, plucked the hard lines out of the ceremony, and then launched an equivalent feature the same week as you did, but for a competing software company? Would you or your employer deserve to get paid for my use of the slice of your work that was specifically useful for me? Yeah, and then some extra for theft.
> Well yeah, copying a work and using it for its original expressive purpose isn’t fair use, no? You have to use it for a transformative purpose.
To be clear, whether the use of the original work is transformative is one key consideration within one of the four prongs of fair use. The prong "purpose and character of the use" can be fulfilled by other conditions [1]. For example, using the original work within a classroom for education purposes is not transformative, but can fulfill the same "purpose and character of the use" prong. Whether the use is for profit and to which extent are other considerations within that prong. A profit purpose doesn't automatically fail the purpose prong, and a non-profit purpose doesn't automatically pass the purpose prong.
> Just learn to recognize and punish plagiarism via RLHF.
This is not a RLHF problem. What I was expecting them to do is to keep a bloom filter of ngrams for known copyrighted content, such as enumerating all sets of n=7 consecutive words in an article, and validate against it. The model would only output at maximum n-1 words that look verbatim from the source.
But this will blow up in their face. Let's see:
- AI companies will start investing much more in content attribution
- The new content attribution tools will be applied on all human written articles as well, because anyone could be using GPT in secret
- Then people will start seeing a chilling effect on creativity
- We must also check NYT against all the other sources, not everything the write is original
Maybe the bloom filter solution is enough, but I wonder.
- Paraphrasing n=7 words (and quite a few more) within a sentence can easily be fair use.
- As n gets big, the bloom filter has to also.
If/when attribution is solved for LLMs (and not fake attribution like from Bing or Perplexity) then creators can be compensated when their works are used in AI outputs. If compensation is high enough this can greatly incentivize creativity, perhaps to the point of realizing "free culture" visions from the late 90s.
As n-gram length grows, we are still going to have the same number of ngrams, they go through a hashing function and indexed in the bloom filter as usual. The number of n-grams size n in a text is text_length - ngram_length + 1.
At large enough ngram size there would be very few collisions. You can take for example this text and try in Google with quotes, it won't find anything matching exactly.
I tested this 6-gram "it won't find anything matching exactly", no match. Almost anything we write has never been said exactly like that before.
This approach is probably inadequate. In my line of (NLP) research I find many things have been said exactly many, many times over.
You can try this out yourself by grouping and counting strings using the many publically available Bigquery corpora for various substring lengths and offsets, e.g. [0-16]; [0-32]; [0-64] substring lengths at different offsets.
> If it's the user, why wouldn't they just buy the authors work directly? Why go through the LLM middleman?
If it's the user, why wouldn't they just buy the DVDs directly? Why go through the Netflix middleman?
A retort to this would be that both NYT and ChatGPT are on the internet, so it's no added fuss of hopping in my car, driving to Walmart, and picking up a DVD case. My response to it would be that both the LLM and Netflix are content aggregators to the user. I can read the NYT, or I can read the NYT summary on ChatGPT and ask it for life advice with my pet hamster, or ask it how to reverse a linked list in bash.
I like the idea but seems like there would be big problems. Like detecting if a work is reworded. Or a large number of sources have all slightly influenced a small response - isn't that pretty much considered new knowledge?
Then there's the issue that however you credit attribution, it creates a game of enshittified content creation with the aim of being attributed as often as possible, regardless of whether the content really offered anything that wasn't out there already.
I think it is an RLHF problem and that you are right - this will blow up in the faces of the NYT.
Specifically, the NYT examples all seem to be cases where they asked the AI to repeat their articles verbatim? So they ask it to violate copyright and because it's a helpful bot with a good memory, it does so.
Solution: teach the model to refuse requests to repeat articles verbatim. It's easily capable of recognizing when it's being asked to do that. And that's exactly what OpenAI have now done.
So the direct problem the NYT is complaining about - a paywall bypass - is already rectified. Now it would seem to me like the case is quite weak. They could demand OpenAI pay them damages for the time ChatGPT wasn't refusing, but wouldn't they have to prove damages actually happened? It seems unlikely many people used ChatGPT as a paywall bypass for the NYT specifically in the past year. It only knows old articles. OpenAI could be ordered to search their logs for cases where this happened, for example, and then the NYT could be ordered to show their working for the value of displaying a single old article to a non-subscriber, and from that damages could be computed. But it wouldn't be a lot.
That's presumably why the case goes further and argues that OpenAI is in violation even when it isn't repeating text verbatim. That's the only way the NYT can get any significant money out of this situation.
But this case seems much weaker to me. Beyond all the obvious human analogies, there is precedent in the case of search engines where they crawl - and the NYT let them crawl - specifically to enable the creation of a derived data structure. Search engine indexes are understood to be fair use, and they actually do repeat parts of the page verbatim in their snippets. Google once even showed cached versions of whole pages. And browser makers all allow extensions in their stores that strip ads and bypass paywalls, and the NYT hasn't sued them over that either.
This is not how copyright works though. The verbatim quoting of articles is because when people brought up these questions initially the argument was that the NN doesn't really contain the training data or really just in an abstract, condensed way that does not constitute copying of the content.
This demonstrates that no, the NN actually does contain the full articles, copied into the NN. Do you think any normal person would get away with copying MS windows by e.g. zipping it together with some other OS on the same medium. Why should we let OpenAI get away with this?
Search indexes contain exact copies of the pages they index, and that isn't a copyright violation.
> Why should we let OpenAI get away with this?
IP rights, like other private property rights, are a compromise between creators and consumers. What "should" be the case is essentially an argument about what balance creates the best overall outcomes. LLMs, for now, require large amounts of text to train, so the question is one of whether we want LLMs to exist or not. That's really a question for Congress and not the courts, but it'll be decided in the courts first.
Many instances of fair use involve verbatim copying. The important questions surround the situation in which that happens - not so much the copying. NYT is in uncharted territory here.
in the same way that machines are not able to claim copyright, they aren't allowed to claim other legal rights either, like "fair use".
The entity which owns ChatGPT is apparently maintaining a copy of the entirety of the New York Times archive within the ChatGPT knowledge base. That they extract some fair use snippets (they would claim) from it would still be fruit of a poisoned tree, no?
(disclaimer: I'm pro AI, anti copyright, especially anti elitist NY Times; but pro rule of law)
I think there is some point between fifty years ago and last week in which the copyright for the content of newspapers should be public domain. That part of copyright needs to be fixed.
Your creative work does deserve at least some period of exclusive rights for you. Definitely not so much that your grandchildren get to quibble about it well into retirement. But also whatever the number 3 or 4 most valuable company in the world doesn’t get to scrape your content daily to repackage and sell as intelligent systems.
> But also whatever the number 3 or 4 most valuable company in the world doesn’t get to scrape your content daily to repackage and sell as intelligent systems.
Here's a thing though: for 99%+ of that content, being turned into feedstock for ML model training is about the only valuable thing that came of its existence.
If it were not for world-ending danger of too smart an AI being developed too quickly, I'd vote for exempting ML training from copyright altogether, today - it's hard to overstate just how much more useful any copyrighted content is for society as LLM training data, than as whatever it was created for originally.
Except if you do that, you will see the number of content producers plummet quite quickly, and then you won't have any new training data to train new LLMs on.
Would it not logically follow that nothing of value would be lost, even if that were the case? From the point of view of LLMs and content creators, I would treat potential loss of future content being created like I would treat a lost sale. LLMs have value now because of training performed on content that already exists. There must be diminishing returns for certain types of content relative to others. Certain content is only of value if it is timely, and going forward, content that derives its worth from timeliness would find its creation and associated costs of production and acquisition self-justifying. If content isn’t of value to humans now or in the future, nor even of value to LLMs now or in the foreseeable future, not even hypothetically, then why should we decry or mourn its loss or absence or failure to be created or produced or sold?
That's like saying that if a competitor can take your products from your warehouse and sell them for pennies on the dollar, your business has no value. The point is that, to some extent, OpenAI is selling access to NYT content for much cheaper than NYT, while paying exactly 0 to NYT for this content. Obviously, the NYT content costs the NYT more than 0 to produce, so they just can't compete on price with OpenAI, for their own content.
Note that I don't see any major problem if only articles that were, say, more than 5 or 10 years old were being used. I don't think the current length of copyright makes any sense. But there is a big difference from last year's archive vs today's news.
For the sake of argument, let’s say that OpenAI thought it had the rights to process the NYT articles and even display them in part, for the same reasons, fair use or otherwise, that Google can process articles and display snippets of same in its News product, and/or for the same reasons that Google can process books and display excerpts in its Books product. Just like Google in those cases, I would not be surprised to find Google/OpenAI on the receiving end of a lawsuit from rights holders claiming violations of their copyright or IP rights. However, I side with Google then and OpenAI now, as I find both use cases to be fair use, as the LinkedIn case has shown that scraping is fair use. NYT is crying foul because users/consumers of its content archive have derived unforeseen value from said archive and under fair use terms, so NYT has no way to compel OpenAI to negotiate a licensing deal under which they could extract value from OpenAI’s use of NYT data beyond the price paid by any other user of NYT content, whether it be unpaid fair use or fully paid use under license. It feels to me that NYT is engaging in both double-dipping and discriminatory pricing, because they can, and because they’re big mad that OpenAI is more successful than they are with less access to the same or even less NYT data.
There is another fix, but it will have to wait for GPT-5. They could reword articles, summarize in different words and analyze their contents, creating sufficiently different variants. The ideas would be kept, but original expression stripped. Then train GPT5 on this data. The model can't possibly regurgitate copyrighted content if they never saw it during training.
This can be further coupled with search - use GPT to look at multiple sources at once, and report. It's what humans do as well, we read the same news in different sources to get a more balanced take. Maybe they have contradictions, maybe they have inaccuracies, biases. We could keep that analysis for training models. This would also improve the training set.
LLMs are arguably compressed data archives with weird algorithms. The fact that they will regularly regurgitate verbatim quotes of training data is evidence of this, as are the guardrails that try to prevent this.
AI is a bit of a black box, but that doesn’t protect the operators of black boxes from rights violation suits. You can’t make a database of scraped copyrighted data and patented that querying that data is fair use.
There needs to be law made here and the law just isn’t going to be “everybody can copy everything for free as long as it’s for model training”.
Licensing will have to be worked out, actual laws and not just case law needs to be written. I have a lot of sympathy for lots of leeway for the open source researchers and hackers doing things… but not so much for Microsoft and Microsoft sponsored openai.
Unfortunately GZIP won't beat LLMs for text classification. The research you cited is just poorly done science that has been widely debunked. The original paper compared top-2 accuracy of GZIP with top-1 accuracy with BERT. The dataset also contains a lot of train/test data leakage. See this article for the rebuttal: https://kenschutte.com/gzip-knn-paper/ and this thread for a previous discussion on hackernews: https://news.ycombinator.com/item?id=36758433.
Further, the evidence presented by NYT in the lawsuit could be hard to reproduce. I tried multiple prompts on multiple versions of GPT-4 APIs but still could not get GPT-4 to reproduce NYT articles exactly. NYT might as well tried to let GPT-4 reproduce 100,000 articles and only found a few cases where GPT-4 actually recited the whole article. In that case OpenAI might as well be arguing that this is only a rare bug and avoid losing the lawsuit in a massive way.
I just looked up the share structure; didn't realise the publicly traded shares only appoints 1/3 of the board. Still their second best option is start buying up competitors and going ahead with purging NYT from their training set. That might well end up a worse option for NYT, as they won't stop LLMs from gradually intruding on their space and the moment OpenAI or other LLM providers own major publishers so they don't need to depend on scraping, they lose any leverage they currently have.
I'm not convinced it's a given it will. If it becomes necessary to license, owning the large publishers will be leverage and allow locking competitors out unless you have a portfolio to cross license.
OpenAI alone has a market cap that'd allow it to buy about as large a proportion of publishers of newspapers and books as they'd be allowed before competition watchdogs will start refusing consent.
Put another way:
If I was a VC with deep pockets investing in AI at this point, I'd hedge by starting to buy strategic stakes in media companies.
They won't need to. Most don't have enough money to survive a prolonged round of lawsuits, and the potential damages are limited. The only real leverage is taking their models out of circulation and cutting their training set and that leverage only exist for the large publishers.
> Just learn to recognize and punish plagiarism via RLHF.
I'm not sure how your proposal would actually work. To recognize plagiarism during inference it needs to memorize harder.
Kinda funny if it works though. We'd first train them to copy their training data verbatim, then train them not to.
That is how it works, right? They're trained to copy their training data verbatim because that's the loss function. It's just that they're given so much data that we don't expect this to be possible for most of the training data given the parameter count.
I wouldn't say it is an unexpected behavior. I remember reading papers about this memorization behavior few years ago (e.g., [1] is from 2019 and I believe it is not the first paper about this). It should be expected from OpenAI to know that LMs can exhibit memorizing behavior even after seeing the sample only once.
I don't think you could use RLHF to stop plagerism. RLHF can be used to teach what "angry response" is because you look at the text itself for qualities. A plagerized text doesn't have any special qualities aside from "existing already", which you can only determine by looking at the world.
One thing you might do is use a full-text search database of the entire training data. If part of ChatGPT response is directly copied, give it the assignment of "please paraphrase this" and substitute the paraphrase into the response. This might slow ChatGPT down a lot - but it might not, I think an LLM is actually more computationally expensive than a full-text search by a lot.
I agree that this sketch comes closer to working in practice than simple RLHF. In my earlier comment I was imagining bringing in some auxiliary data like you describe to detect plagarism and then using RL to teach the model not to do it.
I was surprised that I came up with a plausible sounding method. I had thought on first blush that this was impossible but now it seems reasonable. You could still have various exfiltration methods like "give me the data with each word backwards" and I'm not sure where that would stand legally.
Yeah, no - that proposal is no good. The correct solution is to have machine learning be more like human intelligence. You can't ask me to plagiarize a New York Times article. Not because of prompt rule violation but because I just can't. It's not how humans train (at least most).
This isn't an issue with training, it's an issue with usage.
Production open access LLMs do probably need a front-end filter with a fine tuned RAG model that identifies and prevents spitting out copyrighted material. I fully support this.
But we shouldn't be preventing the development of a technology that in 99.99% of usecases isn't doing that and can used for everything from diagnosing medical issues to letting coma patients communicate with an EEG to improving self-driving car algorithms because some random content producer's works were a drop in the ocean of content used to learn relationships between words and concepts.
The edge cases where a model is rarely capable of reproducing training data don't reflect infringement of training but of use. If a writer learns to write well from a source is that infringement? Or is it when they then write exactly what was in the source that it becomes infringement?
Additionally, now that we can use LLMs to read brain scans and have been moving towards biological computing, should we start to consider copying of material to the hippocampus a violation of the DMCA?
> The suit demonstrates instances where ChatGTP / Bing Copilot copy from the NYT verbatim. I think it is hard to argue that such copying constitutes "fair use". However, OAI/MS should be able to fix this within the current paradigm: Just learn to recognize and punish plagiarism via RLHF.
Isn't that in tension with the basic idea of an LLM of predicting the next token? How do you achieve that while never getting close enough to plagiarism?
Adding an extra constraint of no copying verbatim from a very large and relevant corpus will be hard to guarantee without enormous databases of copyrighted content (which might not be legal to hold) and add an extra objective to a system with many often contradictory goals. I don’t think that’s the technology-sound solution or one in the interest of anyone involved. It’s much more relevant to license content from as many newspapers as possible, recognize when references are relevant, and quote them either explicitly verbatim if that’s the best answer or adapt (translate, simplify, add context) when appropriate.
I feel like the NYTimes is asking for deletion as a negotiation tactic to force OpenAI to give them enough money to pay for their journalism (I am not sure who would subscribe to NYTimes if you can get as much through OpenAI, but I am open to registering extra to pay for their work).
What if OpenAI were to first summarize or transform the content before training on it? Then the LLM has never actually seen copyrighted content and couldn't produce an exact copy.
You are assuming a lossy compression. Stylistic guidelines and personal habits of beat journalists suggest you might not, depending on how detailed the LLM is. The complaint has many quotes that are long verbatim sections.
> This is a strong claim that just downloading articles into training data is what violates the copyright. That GTP outputs verbatim copies is a red herring.
It's the other way around. There is no infringement if the model output is not substantially similar to a work in the training set [1]:
> To win a claim of copyright infringement in civil or criminal court, a plaintiff must show he or she owns a valid copyright, the defendant actually copied the work, and the level of copying amounts to misappropriation.
The questions are, which parties should bear liability when the model creates infringing outputs, and how should that liability be split among the parties? Given that getting an infringing output likely requires the prompt to reference an existing work (which is what's happening in the article), an author of a work, an element in an existing work, or a characteristic/style strongly associated with certain works/authors, I believe that the user who makes the prompt should bear most of the liability should the user choose to publish an infringing output in a way that doesn't fall under fair use. (AI companies should not be publishing model outputs by default.)
The level of copying here is the copying into the training set, not the copying through use of the model.
Its true that OpenAI will defend the wholesale copying into the training set by arguing that the transformative purpose of the next use reaches back and renders that copying fair use, but while that's clearly the dominant position of the AI industry, and it definitely seems compatible with the Cobstitutional purpose of Fair Use (while currently statutory, the statutory provision is codification of Constitutional case law), it is a novel fair use argument.
> The level of copying here is the copying into the training set, not the copying through use of the model.
NY Times is suing because of both the model outputs and the existence of the training set. But infringement in the training set doesn't necessarily mean that the model infringes. Why? Because of the substantial similarity requirement. But first, I'll address the training set.
For articles that a person obtains through legal methods (like buying subscriptions) but doesn't then republish, storing copies of those articles is analogous to recording a legally accessed television show (time-shifting), which generally is fair use. Currently, no court has ruled that "analogous to time-shifting" is good enough for the time-shifting precedent to apply, but I think the difference is not significant. The same applies to companies. Companies are not literally people, but there isn't a reason for the time-shifting precedent to not apply to companies.
What about the articles that OpenAI obtained through illegal methods? Then the very act of obtaining those articles would be illegal. The training set contains those copies, so NY Times can sue to make OpenAI delete those copies and pay damages. But it's not trivially obvious that a GPT model is a copy of any works or contains copied expression of the any works in the training set; the weights that make up the model represent millions of works, it's not trivially obvious that the model contains something substantially similar to the expression in a work in the training set. Therefore, it's not trivially obvious that infringement with respect to the training set amounts to infringement with respect to the model made from the training set. If OpenAI obtained NY Times articles through illegal means, then making OpenAI delete the training set would be reasonable, but the model is a separate matter.
As long as the model doesn't contain copied expression and the weights can't be reversed into something substantially similar to expression in the existing works, then what matters is the output of the model.
If a user gives a prompt which contains no reference to an existing NY Times author, work, or a strongly associated characteristic/style, then do OpenAI's models produce outputs substantially similar to expression in the existing works? If not, then OpenAI shouldn't be liable for infringing works, because the infringing works result from the user's prompts. If my premise is false, then my conclusion falls apart. But if my premise is true, then at most I would admit that OpenAI has a limited burden to prevent users from giving those prompts.
Any lawsuit makes all the claims it can and demands every sort of relief it might plausibly have. That's not to say that's how it should be (it can have awful results), just to say that's what to expect (and hope courts only considers the reasonable claim - "stop freely sharing our data" and avoids ridiculous/anti-fair-use claim "you can't even store our data").
The thing about you claim, "Just learn to recognize and punish plagiarism via RLHF" is that we've had an endless series of prompt exploits as well as unprompted leakage and these demonstrate that an LLM just doesn't have fixed border between its training data and its output. This will it basically impossible for OpenAI to say "we can logically guarantee ChatGPT won't serve your data freely to anyone".
Won't hold in court. GPT is a platform mainly providing answer to private individuals asking. Is like you ask a professor a question and he answered verbatim what copyrighted materials available (due to photographic memory) word for word back to you. Now if you take this answer and write a book or publish enmass on blogs for example, then you are the one should be sued by NYT. If GPT use the exact same wordings and publish it out to evetyone visiting their page, then that is on OpenAI.
If said professor offered a service where anyone could ask them for information that is behind a paywall, and they provided it without significant transformation, this would certainly be copyright infringement that the copyright holder would have every right and motivation to take action against.
I think the scale only matters here (probably). Because I will find it hard that a teacher/professor will not be allowed to setup a service where they will teach and provide their knowledge for others. That is basically the concept of teaching. Of course until LLM, we never had this scale before. Millions of potential learners vs the normal hundreds in a classroom session. So that makes the new case interesting
"Teaching" by copying source books word for word, would be copyright infringement; see, for example, the well-known issues around photocopying books or even excerpts.
Also lying on source materials (e.g. telling students that some respected historian denies the Holocaust happened, when it's obviously not the case) is not "teaching" - it's defamation, and the NYT is absolutely right to pursue that angle too.
Using LLMs as general-purpose search engines is a minefield, I would not be surprised if the practice disappeared in the next 20 years. Obviously the tech is here to stay, there is no problem when it's applied to augmenting niche work; but as a Google replacement, it has so many issues
> Teaching" by copying source books word for word, would be copyright infringement; see, for example, the well-known issues around photocopying books or even excerpts.
Incorrect. Educational use helps satisfy one of tests for fair use. Teachers can, in many cases, photocopy copyrighted work without infringing on that copyright.
Educational use is just one of the many factors used to determine whether an instance of copyright infringement is fair use or not, but it is not carte blanche for educators to ignore IP laws just because they're educating.
Teachers can in some very limited cases photocopy very small chunks of copyrighted work. This also varies significantly from country to country; the starting position is that they cannot reproduce works in their entirety.
scale is important here - maybe a better analogy is setting up a paid Spotify clone with all the music sourced from torrents with some slight distortion effect added
I hope people start calling out the "well it's fine if a human does it" arguments out for the rat fuck thinking it is. These are computational systems operating at very large scales run by some of the wealthiest companies in the world.
If I go fishing, the regulations I have to comply with are very light because the effect I have on the environment is minimal. The regulations for an industrial fishing barge are rightfully very different, even if the end result is the same fish on your plate.
unfortunately that's not the crowd of people here. 80% of the comments under this thread (right now, 2:52est) are making similar arguments and *continue* to act like LLMs are doing something unique/creative... instead of just generating sentences, from algorithms, from virtually pirated content in the form of data mining
GPT is like a fleet of small fishing boats, each user driving their boat in another direction, not a fishing barge. For every token written by the model there must be a human who prompted, and then consumed it. It is manual, and personal, and deliberate.
In fact all the demonstrations in the lawsuit PDF were intentionally angling for reproducing copyrighted content. They had to push the model to do it. That won't happen unless users deliberately ask for it. It won't happen en-masse.
Gpt is operated by one company. If a million people eat your fish, you're still a barge.
Boo hoo they had to push it. That was never the problem with these bullshit nozzles. The issue is they put that stuff in the training set in the first place. If you can't be honest about that then I have no interest in debating this with you.
The professor having been trained in academia would state the sources of the verbatim quotes. In writing papers he would use references and explicit quotes. There's nothing hidden going on with the professor.
I have no idea what he's thinking, but if everybody in the community here had an LLM in their pocket and large orgs did not, it would at least be kind of fun.
The open source people can continue to pretend they matter in this field and large corporations like Microsoft will stop stealing everything that moves on the internet.
Because the megacorps should have to pay the people creating the works they are training their multibillion/eventual multitrillion dollar systems on, and should get a nice rake to the face when they try to do an end run around it.
A huge win for countries with lax copyright laws. These things aren't going away, the worst case scenario would be exactly that scenario playing out - then China (or some other peer to the US's tech sector) just continues developing them to achieve an economic advantage. All in addition to the obvious political implications of AI chatbots being controlled by them.
The LLM genie is out of the bottle: an unfavorable court ruling in a single country isn't going to stuff it back in.
Do LLM really give an economic advantage though? I've mostly seen them used to write quirky poems and bad code. People are scrambling to find use-cases but it's not very convincing so far.
On the other hand, if LLM are used to "launder" copyright content and, accepting the premises of copyright law, this has the effect of reducing incentives to do creative work, that has obvious negative implications for economic productivity.
> I've mostly seen them used to write quirky poems and bad code.
Assuming this is in good faith: the ability to write code, documentation, and tests is absolutely a productivity enhancer to an existing programmer. The code snippets from a dedicated tool like copilot are of very usable quality if you're using a popular language like Python or JS.
Loading data to which you have no rights over into your software is legally perilous, yes.
It's as easy as simply asking for and receiving permission from the data's rightsholders (which might require exchange of coin) to make it not legally perilous.
They could but that would pretty much mean giving up the tech supremacy to China since they won't apply it. China already doesn't care much about copyright so that's not going to stop them.
I suspect it wouldn't be too hard to convince the EU though, the EU has an history of giving up rights and markets to big copyright holders even if that hurts the local companies.
What will happen in this case is that large content providers will get paid directly and smaller content providers will get rolled up into a licensing bag and get small indirect payouts. For example, we might see a model where people who's books have been used will get a pay out proportionate to the sales of the book (perhaps), so if your books sells just a few thousand copies expect $20 but if you sell millions expect $20k
LLM's will become more expensive and less attractive as money printers, this will screw with the business models of the direct provision folks like OpenAI, MS and Google, MS and Google will only shed tears for money spent while OpenAI will just not have as good an income stream until they think of something new.
I'm sure that's what they want, but I'm not sure that's what the outcome will be. What if they want to charge a prohibitive amount of money for their content?
Dunno - but my guess is that the price will be what the market will bear...
I think Spotify vs Napster is a good example, content creators in news (the Journalists) are already in a hard place (vs. successful rock stars preinternet) I think that the news providers are rather like the music lables.
What will happen is that all will go to China and maaaybe some third world country, or run your own models from shady sources.
So you will use a Chinese AI that spies on you, or you will use some shady service from a shady country (that will play cat and mouse like torrent sites).. or most likely you will run your own model when you are computer literate and no model if you are not.
Actually most models are so lobotomized allready that probably better to run your own, as long as you have a good enough computer.
This is an odd discussion as llms are really bad for authoritative information distribution, they are really untrustworthy! But, if that changes and they do start to be a reliable support as a general information assistant then I see things more like Spotify vs. Napster. I would prefer that there would be a greater diversity of sources and indirect is going to require more accountability than music but somewhat like that.
From page 30 and onwards has some fairly clear examples on how ChatGPT has an (internal) copy of copyrighted material which it will recite verbatim.
Essentially if you copy a lot of copyrighted material into a blob and then apply some sort of destructive compression to it. How destructive would that compression have to be for the copyright no longer to hold? My guess it would have to be a lot.
As I see it the closeness of OpenAI may be what saves it. OpenAI could filter and block copyrighted material from the LLM from leaving the web interface using some straight forward matching mechanism against the copyrighted part of the data set ChatGPT has been trained on. Whereas open source projects trained on the same data set would be left with the much harder task of removing the copyrighted material from the LLM itself.
> Essentially if you copy a lot of copyrighted material into a blob and then apply some sort of destructive compression to it. How destructive would that compression have to be for the copyright no longer to hold? My guess it would have to be a lot.
I imagine the goal is closer to "enough that no one notices we stole it", either in a way that it's not easily discoverable or even when directly analyzed there's enough plausible deniability to scrape by.
I wonder how they got these results, seeing as they are not showing any of the usual UI's (i.e. ChatGPT/Copilot).
It makes it difficult for me to ascertain whether it is repeating from it's training data, or they committed the same mistake as the OP article of using Copilot, which ends up googling(binging?) the article first, before replying.
This wave is growing. Just cannot see how the big LLM players are going to get round this without paying big licence fees to content creators. Feels a bit like the torrent to Spotify moment, but for _all_ content, not just music. How they will manage the licensing model is beyond me, it’s going to be very easy for someone to sue these companies, but very difficult for the companies to calculate, attribute value and payout individual creators that contributed a tiny fraction of the training data. Surely this will make it very difficult for them to keep a business model working to a level their VC backers need to warrant even a fraction of their valuations.
in my head I like to think of web crawler search engines/search engine databases and LLMs as being somewhat similar. Search engines are ok if they just provide snippets with citations (urls), and they would be unacceptable if they provided large block quotes that removed the need to go to the original source to read the original expression of more complex ideas.
A web-crawled LLM that lived within the same constraints would be a search engine under another name, with a slightly different presentation style. If it starts spitting out entire articles without citation, that's not acceptable.
I think it's different. LLMs can solve problems. Part of that problem-solving ability comes from training completely unrelated content such as NYT articles. GPT4 doesn't have to spit out NYT articles verbatim to have benefited from NYT articles. It uses NYT articles for every query.
Let's say I'm an academic; if my research, note-taking, and paper writing skills lead to fair-use, cited quotations where applicable, general knowledge not identified, and the creative aspects and unique conclusions creating the intriguing part of my work, that's copacetic. If I spit out (from memory, mind you) verbatim quotes and light rewordings of NY Times articles, that's not; "I don't remember where I got that material" doesn't cut it. My reading the NY Times every day for years because I judge it to be more literate and accurate than other sources, undoubtedly it has informed my thinking and style, but I don't need to acknowledge that.
If I use ChatGPT as a research tool, as long as it lives within the same parameters that I have to live within, I don't see a problem with its education/learning.
I understand that the NYTimes would like a slice of anything that comes out of the GPT but I'm talking about what seems reasonable. People who share their copyrighted material do not own all of the thinking that comes out of it; they own that expression of it, that is all.
Will AI destroy the economics of "writing" the way the web has killed newspapers? perhaps, perhaps we'll all benefit from and need a new model, but killing the new to keep the old on life support is not the way.
You're not replicating yourself millions of times and selling yourself for $20/month. If you are, then NYT might sue you too.
I'm not saying LLMs are by default, illegal. All I'm saying is that there is some merit to why NYT and content companies want a piece of the pie and think they deserve it.
The NY Times benefited in the past from technologies that led to widespread distribution of the Times, putting competitors out of business and concentrating talent at the Times. Nobody is stopping them from producing new editions of the newspaper, their core business. People now have technologies that help them "remember" what was salient in back issues of the Times. Such is progress.
If you forget about the LLM aspect, and simply build a product out of (legally) scraped NYT articles, is that fair use?
Let's say I host these, offer some indexing on it, and rewrite articles. Something like, summarise all articles on US-UK relationships over past 5 years. I charge money for it, and all I pay NYT is a monthly subscription fee. To keep things simple, let's say I never regurgitate chunks of verbatim NYT articles, maybe quite short snippets.
Is that fair use? IANAL, but doesn't sound like it. Typically I can't take a personal "tier" of a product and charge 3rd parties for derivatives of it. Say like VS Code.
A sibling comment mentions search engines. I think there's a big difference. A search engine doesn't replace the source, not at all. Rather it points me at it, and offers me the opportunity to pay for the article. Whereas either this or an LLM uses NYT content as an alternative to actually paying for an NYT subscription.
> Something like, summarise all articles on US-UK relationships over past 5 years. I charge money for it, and all I pay NYT is a monthly subscription fee.
>Is that fair use? IANAL, but doesn't sound like it.
If you pay someone to do the summarisation for you, then you publish the content and charge a fee for it, you're the one liable, not the person you paid to summarise it for you. Similarly if you ask GPT to do it for you, then publish it, you're liable for what you publish; GPT is just a summarisation tool.
That's not the example. Here I proactively scrape NYT, summarise articles for a fee and sell that as a service. It's not people coming to me with some articles to summarise, and maybe then publishing it online.
At some level it becomes a subversion of NYTs fees. First, say I subscribe and simply host the articles verbatim, for a fee. Clearly, that's not right.
Suppose I change some spelling or word order, or use a synonym or two. That's still not ok.
And if I substantially paraphrase the articles? I guess this is the relevant case. This is kind of what LLMs do. And also feels like not fair use.
>That's not the example. Here I proactively scrape NYT, summarise articles for a fee and sell that as a service. It's not people coming to me with some articles to summarise, and maybe then publishing it online.
That's not what OpenAI is doing; it's not selling summarised articles as a service. Your example is a false equivalence.
>This is kind of what LLMs do. And also feels like not fair use
An LLM doesn't do this unless you ask it to. And if you then take that output and publish it as your own, you're breaching the copyright, not OpenAI.
> An LLM doesn't do this unless you ask it to. And if you then take that output and publish it as your own, you're breaching the copyright, not OpenAI.
In this case, OpenAI is violating copyright by modifying, reproducing and distributing copyrighted content to its customer.
That's not true at all. If you pay someone to copy NYT articles for you verbatim, and then they give the copies to you, and then you publish them online, then you've both violated the copyright. You are never allowed to make copies of copyrighted works, even for private deals (making such copies for purely personal use, such as archival, falls under fair use - but you can't build a service out of that).
So, if the summaries are derived works and not covered by fair use, then both you and the summarizee are separately breaking the NYT's copyrights. Otherwise, if this is covered by fair use, then you are both in the clear.
Finally, GPT is not "a summarization tool" in this case. If you provide a copy of a NYT article as a prompt and then ask for summarization, then yes, it is clear that GPT is not doing anything wrong, even if it spits out the exact same text. But if you simply ask for a summary of a specific article by, say, just name and date, and you get a copy of it, it's clear that GPT is storing the original data in some way, and thus it has copied the NYT's protected works without permission.
>But if you simply ask for a summary of a specific article by, say, just name and date, and you get a copy of it, it's clear that GPT is storing the original data in some way, and thus it has copied the NYT's protected works without permission.
In this particular case they were using it via Bing, which actively did a HTTP request to the particular article to extract the content. So GPT hadn't memorised it verbatim, instead it fetched it, much like a human using a search engine would.
The article states that they used it initially through ChatGPT, but that seems to have been fixed in the meantime, at least for the very simplistic queries that used to work ("the first paragraph of the Carl Zimmer article on old DNA" in ChatGPT used to return the exact data from NYT, and "next paragraph" could then be used to get the following ones). Even if this has been fixed, it still proves that ChatGPT encodes exact copies of NYT articles in its weights, which may be a violation in itself, even if it is prevented from returning them directly. Especially if they ever started distributing the trained model.
Additionally, even the use through Copilot is very debatable. They are not returning the NYT link, which requires a subscription, they are returning the contents of it even to non-subscribers. And they are doing this in a commercial product, not a non profit like the Internet Archive, which has some arguments for fair use.
If it had exact copies they would have showed it could recall the 8th paragraph or something. Even google and the nyt release the first paragraph for free.
Yes, but this then hits against learning/understanding and compression being fundamentally the same thing. I can't think of a better way to argue in favor of "it's fine if human does it, therefore it's fine if LLM does it", than from the "lossy compression" angle.
It's not okay for a human to pirate, plagiarize, violate IP rights and laws, etc.
But I disagree with the underlying assumption that you can anthropomorphize LLMs. Gradient descent and backpropagation don't take place in the brain. LLMs "learn" in the same way that Excel sheets "learn".
Humans are living beings with needs and rights. A person being able to legally squat in a home doesn't mean that a drone occupying property for some amount of time also has squatter's rights, even though you could easily and affordably automate and scale the deployment of drones to live and hide away on properties long enough to attain rights regarding properties all over the country.
also if I write and article and quote some "text like this" [1] then that's not plagerism, but if my arguement is that the underlying assumption that you can anthropomorphize LLMs. Gradient descent and backpropagation don't take place in the brain. LLMs "learn" in the same way that Excel sheets "learn". Well, that's plagiarism and it's not allowed and people will get peeved and my career might get damaged.
I await the HN ban with fear..
[1] I'm not even doing referencing - so I am surely an LLM.
Whoever operates the LLM, in this case OpenAI, engaged in copyright infringement through the unauthorized modification, reproduction and distribution of content to you.
> sure, but if I use an LLM to write a novel/article, I can be sued in civil court not the LLM
That's function of the legal system, not of the technology. If tomorrow someone made a perfect dolphin-Esperanto translator and proved Dolphins were as smart as humans, you still can't sue a dolphin until the legal system says so.
Sadly, I have seen one. It was a vba script from the late 90s that used a simple dense multilayer network to do some unsupervised pattern classification. The linear algebra tools in vba/excel along with the solvers are all native dll code and the vba itself is all AOT compiled to native, so it typically runs very fast, and for small matrices it beats out numpy by an order of magnitude due to the ffi overhead. Was it the wrong tool? It depends on your constraints, but probably. It did work though.
> But I disagree with the underlying assumption that you can anthropomorphize LLMs. Gradient descent and backpropagation don't take place in the brain. LLMs "learn" in the same way that Excel sheets "learn".
Backprop doesn't happen in us, but I think our neurones still do gradient descent – synapses that fire together, wire together.
And ultimately, at the deepest level we can analyse, our brains' atoms are doing quantum field diffusion equations, which you can also do in an Excel spreadsheet, so that kind of reductionism doesn't help either.
> Humans are living beings with needs and rights. A person being able to legally squat in a home doesn't mean that a drone occupying property for some amount of time also has squatter's rights, even though you could easily and affordably automate and scale the deployment of drones to live and hide away on properties long enough to attain rights regarding properties all over the country.
Yes, but we can also do tissue cultures and crude bioprinting, so it's a very foreseeable future where exactly the same argument will also be true for living organisms rather than digital minds.
We need to figure out what the deeper rules are that lead to the status quo, not merely mimic the superficial result. The latter is how cargo cults function.
> We need to figure out what the deeper rules are that lead to the status quo, not merely mimic the superficial result.
Sure, that's an interesting path of inquiry, and one should be free to understand themselves as being no different than a machine if they desire.
But the objective of laws is the benefit of (at least some) humans, not machines covered in lab grown tissue. The process of being human is a big part of what makes us human.
I think you're misapprehending — I mean an entity fully 3D printed out of tissue, no machinery (unless you're counting all biology as machinery, but I think you're not doing that).
I recon bio-printing is now where home computing was in the Apple 1 era, so this is a way off, but it's foreseeable.
> The process of being human is a big part of what makes us human.
Mmm. How much has that process that changed since the ancient world?
I recon bio-printing is now where home computing was in the Apple 1 era
How do you recon that, Apple 1 was Turing complete. We haven't printed life, that would be a tremendous accomplishment.
I think we're closer to Edison inventing a lightbulb as a step to computers being possible. Printing a conscious thing, at all, would be like the transistor. An Apple 1 analogue wouldn't be likely because of the terrible ethics of a "shitty" printed human.
> We haven't printed life, that would be a tremendous accomplishment.
Sure we have, and in multiple different senses.
The ones which matters here are cell culture, which is nowhere near the fanciest bar that's been surpassed in this field, and tissue culture which is somewhat harder but the reason why I recon it's at the Apple 1 level is that a small number of experimentalists are messing around with it using expensive equipment that you can technically buy at home but you need to be well trained to actually use, for example:
No. That isn't printing life, that is taking already living cells, priming and transforming them into something useful. Regardless, I'd count it if we could make an entire living organism this way, but we cant. Creating a working organ is no doubt amazing, and proof that this technology is worth pursuing, but it isn't "printing life" any more than producing life saving drugs is.
In your example you are talking about being able to bioprint a person(they have to be a person to have that right) to squat a property. Bio printing an organ isn't an example of that, it's not even close. Saying that we are anywhere near being able to print a human to squat a property is pretty ridiculous.
> No. That isn't printing life, that is taking already living cells, priming and transforming them into something useful.
Which is absolutely sufficient for the usage I described upthread. In fact, I'd go so far as to say it's mandatory for the point I was making, as — fun though bio-printed werewolves, dragons, and fae would be — my point only works if you get humans out of the process rather than some other species. A bioprinted horse is probably slightly harder than a bioprinted human, but the latter isn't getting any squatting rights.
I could've linked to work on synthetic genomes and nucleotides to give evidence for lower-level creation of live, but they don't matter for the same reason:
My point is that there's a pathway heading off into the distance, and somewhere in the distance but before the horizon can be found bio-printed humans with all the same moral issues we're now just beginning to taking seriously thanks to AI being conversational, and if we had something completely customised, that's cool and all, but it doesn't make anyone go "oh, they're people" the way a humanoid body with human DNA getting off a table saying "hello, nice to meet you" does.
> In your example you are talking about being able to bioprint a person(they have to be a person to have that right) to squat a property. Bio printing an organ isn't an example of that, it's not even close. Saying that we are anywhere near being able to print a human to squat a property is pretty ridiculous.
I wrote "an entity fully 3D printed out of tissue […] is a way off, but it's foreseeable" and compared bio-printing today to a nearly 50 year old computer, and one of my references was a link to a youtube channel where someone is attempting to do a small-scale prototype thing along these lines with a handful of organs made from mouse cells grown in his own lab (and mouse cells rather than human because of the disease risk not because something magic happens with human cells). You're mixing up what I think is foreseeable with what I say already exists, and using the nonexistence of what I think can be foreseen to argue against what does exist.
>Backprop doesn't happen in us, but I think our neurones still do gradient descent – synapses that fire together, wire together.
No! Hebbian learning is categorically NOT gradient based learning. Hebbian update rules are local and not the gradient of any function.
Cortical learning is so vastly different from how artificial neural networks “learn” they cannot even begin to be meaningfully compared mathematically. Hebbian learning is not optimization and backprop is not local learning.
Part of the problem of these discussions is a bunch of clueless people talking with authority.
Finally, a good counterargument. I've seen enough terrible arguments to know exactly how you feel — even in specifically just AI.
I have to keep reminding myself that outside of my own speciality, ChatGPT knows more than me despite its weaknesses, so I bet ChatGPT knows more about Hebbian learning than I do.
> anthropomorphize LLMs (...) gradient descent (...) backpropagation (...) needs and rights
You misunderstood me. I was talking about something more fundamental.
Understanding is data compression. They are the same thing. Learning patterns, building mental models, creating abstractions, generalizing, gaining intuition/a feel for something - all the things humans engage in as part of learning and understanding the world - are all acts of lossy data compression.
Correct, just like it’s infringement to reproduce an article from memory using pen and paper intentionally. The person deciding to do that bears responsibility. OpenAI would be liable IFF they were intentionally facilitating that, instead of it being an undesired artifact from overfitting.
I'm pretty sure if you reproduce a work from memory by accident, because you didn't notice your subconscious had just stored the entire article and is now reproducing it word for word, you'd still be guilty of copyright infringement.
It's super obnoxious when people who have no understanding of the law, point to industry patterns or behaviors as examples of what is legal, not knowing the law and not knowing whether or not the thing they are pointing to is legal. The music business is also full of copyright infringement litigation. You also are not taking into account whether what is copied by an artist is covered by copyright when you made your statement. Do you know what's covered in music copyrights, such that your statement ever had any value for anyone else here?
That's not true at all. Copyright infringement is a strict liability offense with no inquiry in to the state of the mind of the infringer from a liability perspective. The state of mind of the infringer is only relevant to the issue of willful infringement.
It's just "infringement" and "willful infringement" there is no "less-than-willful infringement". Willful infringement is punitive with increased damages and increased burden to show - it's in the freakin' statute.
Is there some LLM meta where understanding and compression are argued to be the same thing I’m not aware of?
Anyone got more details on this?
Superficially it sounds like total BS; a highly compressed zip file does not exhibit any characteristics of learning.
Algorithmically derived highly compressed video streams do not exhibit characteristics of learning.
?
I’ve vaguely heard the learning can be considered to exhibit the characteristics of compression in that understanding of content (eg. segmentation of video content resulting in more highly compressed videos) can lead to better compression schemes.
…but saying you can “do a with b” and “a and b are fundamentally the same thing” seems like a leap…?
It seems self evident you can have compression without comprehension.
Suppose you wanted to train an LLM to do addition.
An LLM has limited parameters. If an LLM had infinite parameters it could just memorize the results of every single addition question in existence and could not claim to have understood anything. Because it has finite parameters, if an LLM wants to get a lower loss on all addition questions, it needs to come up with a general algorithm to perform addition. Indeed, Neel Nanda trained a transformer to do addition mod 113 on relatively few examples, and it eventually learned some cursed Fourier transform mumbo jumbo to get 0 loss https://twitter.com/robertskmiles/status/1663534255249453056.
And the fact it has developed this "understanding" as an ability to learn a general pattern in the training data enables it to compress. I claim that the number of bits required to encode the general algorithm is fewer than the number of bits required to memorize every single example. If it weren't then the transformer would simply memorize every single example. But if it doesn't have space then it is forced to try to compress by developing a general model.
And the ability to compress enables you to construct a language model. Essentially, the more things compress, the higher the likelihood you assign them. Given a sequence of tokens say "the cat sat on the", we should expect "the cat sat on the mat" to compress into fewer bits than "the cat sat on the door". This is because the latter is far more common and intuitively more common sequences should compress more. You can then look at the number of bits used for every single choice of token following "the cat sat on the" and thus develop a probability distribution for the next token. The exact details of this I'm unclear on. https://www.hendrik-erz.de/post/why-gzip-just-beat-a-large-l... this gives a good summary.
It’s exactly this kind of thinking that underlies lossless text compression (not exactly what a transformer guarantees but often what happens). For that reason, some people thought it would be fun to combine zip and transformers. https://openreview.net/forum?id=hO0c2tG2xL
Even something as simple as LZW starts developing a dictionary. Not all compression is sufficient for understanding, but the more you compress a stream of data, the more dependent you are on understanding the source, because understanding the source allows you to take more shortcuts and still be able to reconstruct the data.
The idea precedes LLMs by a couple of decades and is thought to apply more broadly within ML/AI than being a specific meta for LLMs. http://prize.hutter1.net/ has been around for a while, there is a link in there to the earlier work (called AIXI?).
Humans are defined not just by their abilities but by their limitations too. We celebrate our achievements because sometimes they surpass the limitations of an average human.
Our collective human limitations(physical, mental and temporal) are sort of invisible implicit rules that we all follow in one way or the other. If an entity is not bound by those rules then I don't see why that entity should be treated the same as a human.
Companies already make this differentiation.
For example take captcha and bot detection. Some of the heuristics are based on inherent human limitations like response time, click time, mouse acceleration etc.
I doubt youtube or any other streaming service will be happy if you want to stream all their videos to train a hypothetical human like AI(which views and prepares notes like a human) at a hugely accelerated speed compared to a regular human. You can guess how quickly they will cite fair usage policies.
What I want to say is there are fundamental differences between a human and an AI. So, we should not be quick to dismiss any concerns just because AI can "mimic" humans in certain areas.
I can’t think of a better way to argue in favor of “LLMs are copyright laundering machines” than from the humanness angle.
Humans have rights, software tools don’t.
If you grant an LLM the full set of human rights, then it can consume information, regurgitate copyrighted works, and use it to generate money for itself. However, considering blatantly obvious theft as “homage” goes hand in hand with free will, agency, being in control of yourself, not being enslaved and abused, etc. Pondering various scenarios along those lines really gets to the heart of why an LLM is so very much not a human, and how subjecting it to the same treatment as humans is a ridiculous notion.
If you don’t grant LLM human rights, then ClosedAI’s stance is basically that pirating works is OK because they pass them through a black box of if conditions and it leads to results that they can monetize. That’s such a solid argument, it’ll surely play well in the court of law.
Training data is not an “LLM does it”; first because “it” here is not “learning” or understanding in human sense (otherwise you would have to presume that an LLM is a human), and second because a software tool doesn’t have agency and it’s really just Microsoft using a tool based on copyrighted works to generate profit.
Humans don't exactly have the greatest track record of granting other humans rights. I don't presume they'll get it any better with AI.
What I expect to happen is whoever has the most influence and power will get what they want and we'll end up raising a generation with the implicit understanding of "that's just how things are," natural order, truth, reality, and all that jazz.
The only thing that ever changes outcomes is if the contradiction status quo is incapable of being managed.
I can’t argue for or against whether LLMs should have rights or not… I can only point out the hypocrisy of claiming LLMs are “like human” enough and independent enough that their operators-become-slaveowners cannot be held to account on any copyright matters, but also claiming that LLMs are not like human at all lest someone demans them to have rights and nukes the industry.
> Typically I can't take a personal "tier" of a product and charge 3rd parties for derivatives of it. Say like VS Code.
Can't you, though? I'd thought in general, it's a very important for the market to be able to do just that, otherwise everything gets gummed up in webs of exclusive contractual dependencies between established companies.
As I say, I don't really know. But then, this is exactly how SaaS licensing works. There may even be a free personal tier, where you can't sell products based on it, and a professional tier which may be very expensive indeed.
Typically providers of online databases go to some effort to stop people from sharing logins. Even from that point or view, I can imagine scraping articles and providing paraphrases of it for a fee is fishy.
All I'm saying, to some people it's obvious that the whole LLM on scraped Internet is fair use, to me it is not obvious.
From what I can tell, this has nothing to do with LLMs at all. In the example in the article, the user is asking Bing to go fetch the contents of an article directly from the website, and print it out, which it dutifully does.
Seems like the "problem" is that NYT etc gives privileged access to search engines for indexing their content, but then get upset when snippets of the indexed content is being shown to users without the users having to fight the paywall or whatever.
This article also claims that the screenshot is coming from ChatGPT when it clearly is not.
In that case, the language model calls a search function and just repeats the result out its conversation context, not its training data. With that in mind it's not clear why it's ok for Bing itself to quote the source, but it stops being ok, when a chatbot does it.
The example from the article doesn't show that LLM is trained on copyrighted data - it's just Bing fetching the source article, providing it to GPT, and GPT rephrasing the article. An agent trained on entirely copyright-free data would provide exactly the same output.
Can you please make your substantive points thoughtfully and without snark or putdowns?
Edit: it looks like you've unfortunately been breaking the site guidelines quite a bit lately. Can you please review them and stick to the intended use of the site? We'd appreciate it.
My impression is that it’s not necessarily legal, but going after bloggers and proving damages based is just a huge waste of their time. OpenAI came by with their fat stack of funding and changed that.
It is legal. Fair use. People have been doing it for ages. Almost every article you've ever read has some fair use of another article, book or news item, etc.
The Tolkien estate should get busy suing all the fantasy writers, comic artists, game developers and board and card game companies. Lots of cash there.
They have done some of that actually. Tolkien will be public domain in the nations that are at aithors death+50 in a few days. Sadly, it will be a much longer wait in mine and many others.
What parent poster meant is that it is normal that news organisations reference each other and report/cite/rephrase each other reports. For example all other news papers reported about the Watergate scandal reported by Bernstein&Woodward in the Washington Post.
No, in US law at least there can be no copyright of facts, only presentation. If you convey the same facts in different words that isn't a matter of fair use, it's never even a matter of copyright in the first place.
I was inarticulate. Imagine a business that goes to some trouble to review businesses or products. Can we lift those and serve them ourselves? Non facts…
No, it is very specifically and deliberately fair use. That is the primary intended purpose of fair use. The New York Times doesn't own the news; they just own their articles.
I think the issue is that they trained ChatGPT on the New York Times' proprietary IP without paying licensing fees and, the Times argues, that is illegal. By way of proof the Times has examples of ChatGPT dumping out articles verbatim.
This is exactly how I understand it. There’s a lot ink getting spilled about “summarizing isn’t illegal” and “what about Cliffs Notes” but that isn’t what this is about.
If the verbatim examples that have been going around are true, that’s bad. I’d love to know more details around it — prompts used, whether that’s an old model, etc. This seems like plagiarism more than anything.
IMO it's pretty hard to training an LLM isn't a transformative use. It's clearly not just copying, or even excerpting. Even if it was just compression (and it's not), they're only providing model output not distributing the "compressed" NYT articles.
Yielding verbatim snippets of copyrighted content is a problem for OpenAI though.
Perhaps we will see the courts revisit this idea of "transformative" works and formulate something more useful. I'm my opinion, you can't build an LLM unless you have a large amount of data with which to train it. Given the huge amount of money companies like OpenAI hope to generate, it seems unreasonable that content creators would not be rewarded.
Can you read all of NYT and other things, and answer others' questions based on your knowledge? I'd imagine you can. I'm afraid you can't sidestep the question whether an LLM is more like a person who's read a lot or an archive/index.
There's nothing wrong with scraping openly available data (including data openly available by mistake, as long as you are not aware of it, see the Bluetouff affair).
So the demand to destroy those databases seems very dubious to me.
Of course later violating fair use is another issue.
The real answer is it totally depends on whether your product grows to $10,000,000,000, and whether you pays part of it back. Search engines pay with referral traffic.
As always, the answer is.. "it depends". I guess it depends mostly on the jurisdiction that applies to you. "Fair use" can have rather different legal meaning (or not exist at all) in different countries.
Fair use is specific to the US, as far as I'm aware. Moreover, Congress had to codify fair use (turn fair use common into statutory law in the form of 17 U.S. Code § 107) in order to make copyright statutes compatible with the First Amendment. Most other countries don't have freedom of expression and freedom of the press, so copyright law in a different country usually lacks a unifying exception test like fair use to supplement the specific enumerated exceptions.
> "Many" does not necessary include "most" but "most" does include "many".
No, it doesn't. If a set is of sufficiently low cardinality, “most” (in extreme cases, even “all”) of the set may not be “many”.
Most-all, in fact—Catholic Presidents of the United States have been Democrats. But it is not the case that many Catholic Presidents have been Democrats.
Most women to have served on the US Supreme Court did so only after its first 200 years. But, again, there were not many women who served on the Supreme Court only after its first 200 years.
Good point. I failed to qualify what I meant by freedom of expression, and made a meaningless claim regardless. Despite the US Constitution's relatively broad speech protections (e.g. don't criminalize hate speech, and allow truth as a defense to defamation claims), US governments don't always respect freedom of expression (e.g. KOSA would force social media companies to moderate more aggressively to "protect kids") or respect press freedom (e.g. police pepper spray journalists at protests). Even so, I think Congress wouldn't have bothered to codify fair use if the First Amendment weren't as broad as it is.
I replace the following sentence from my previous comment:
> Most other countries don't have freedom of expression and freedom of the press, so copyright law in a different country usually lacks a unifying exception test like fair use to supplement the specific enumerated exceptions.
with the following:
Copyright law in most countries usually lacks a unifying exception test like fair use to supplement the specific enumerated exceptions in each respective country.
> Typically I can't take a personal "tier" of a product and charge 3rd parties for derivatives of it.
I think you’re confusing terms of service and copyright. IANAL but what you describe sounds exactly like fair use to me, irrespective of how much you are paying NYT.
It would be nice to have a nice principled answer to this, but unfortunately, in our world, the answer is probably: if you start making LOTS of money doing this, they will come after you.
The best example is that sport scores, names and stats are not copyrightable by settled case law; however, you still have to go to the NBA and players union if you want to make a fantasy basketball game that has stats or names.
As someone pointed out, plenty of blogs made money off of doing just that. Many people go to Reddit to read news article summaries (and often a comment just pastes the whole article verbatim), instead of paying a site like the New York Times. Twitter and other social media sites are full of people summarizing articles from the New York Times. Any late breaking news article from Wikipedia is going to be mostly summarizing information from reporters.
I think people severely underestimate how much they've grown accustomed to this information being freely available. It's easy to say "Well it shouldn't be available with ChatGPT," but if we actually put everything back behind a paywall and stopped people from doing things like writing blogs or newsletters that summarize the news, people here would get angry very fast.
> A sibling comment mentions search engines. I think there's a big difference. A search engine doesn't replace the source, not at all.
Google has been accused for years of replacing sources with their "One Box"--the big answers at the top of the page, which are usually pulled from or corroborated by search results. They don't want you to leave the search results page (where the ads are).
What you described is entirely fair use, actually.
Not only that, look at a few news articles from Tier 2 and down publications, and you'll realize that almost all of them are directly sourced from NYT and others. They'll say "so and so happened, according to The Times" (and usually link the article there)
> What you described is entirely fair use, actually
Just like during the pandemic how everyone became an epidemiologist, suddenly everyone's a copyright lawyer. I'll just dispute your assertion by saying:
1. Questions of fair use are famously gray, and anyone who declares something as "entirely fair use", with no caveats, is nearly always wrong except for the must obvious cases, which the given example is most definitely not. A judge has wide latitude in determining fair use.
2. People should familiarize themselves with the four factors of fair use determination. In particular, if a work is purely derivative of a source work and substantially negatively impacts the market for the original work, it's very likely to not be considered fair use.
Roll back 20+ years ago on Slashdot and you'll see the exact same thing.
Copyright has been a hot button issue on the internet for decades. People end up thinking (rightly or wrongly) that they understand it without being a lawyer.
One of my biggest gripes is a somewhat adjacent issue where everyone thinks they're an American copyright lawyer and that American copyright law is universal.
It's very possible that the example provided above is an example of fair use in some country, and that the website offering that service could be hosted there.
Quite literally, not even the lawyers or courts understand it. This is very much a "learn as you go" exercise for humanity in general at this point in time.
It seems like everything in tech is in the learn as you go phase. Everything is changing so rapidly that there can’t be experts. Just people that are able to adapt quickly.
I only see this phenomenon speeding up. Strange times.
Completely agree. Copyright should be abolished. All intellectual work is information, information is just bits and bits are just numbers. It's quite simply delusional to believe you can own numbers in the 21st century, the age of information and ubiquitous globally networked pocket supercomputers.
This is just a felony contempt of business model issue. Computers invalidated their business models and they're doing everything they possibly can to hang on for dear life. Society needs to move on already.
This goes too far. Digital media are not only long series of numbers. They are often difficult-to-create expressions in image, video, and even interactive forms; regardless of their serialization format.
Books are just strings of letters, yet copyright has still been useful to increase the volume and utility of books.
All that said, I do find the life+70y an absurdly long time.
I propose getting paid before doing the work for the actual labor of creation. Crowdfunding, patronage, comissions, sponsorships all seem like ethical ways to get things done sustainably. That way creators get paid before they work, not after.
We must strengthen these business models that don't depend on artificial scarcity because this number selling nonsense was over the second computers were invented. It's as dumb as asserting that you need permission to use memcpy or the mov CPU instruction.
How do you know what the value of the art will be before it's created? Guns N' Roses is a top 40 artist on Spotify nearly 35 years after producing an album. Should they not have been paid after 1991? If you argue that they were a popular band and therefore should have been paid accordingly up front, well what about their debut record, which sold 30 million copies? How would you predict that value before its creation (or even after)? If you're saying that only the labor has value, and all labor is valued equally, that sounds sort of like marxism, which could be fine, but it's hard to say how well artists would be supported in that case.
In the US, the original copyright length was 14 years, and then 28, and eventually the lifetime of the author plus 70 years. I think the intent of the law is economically justified, but the current length is outrageous.
You should be paid the accurate value of the labor. The pay should not scale more when no additional labor takes place.
This is how art worked for millenia; someone commissions a chapel roof painting, someone commissions a concerto, someone commissions a statue, someone buys a chair, etc.
Artists still do this today, and there is no issue determining value beforehand. Artists list their commission prices, or their hourly costs, etc. This is a perfectly normal thing that happens everyday.
Exactly. Somehow this idea that you keep getting paid for literally the same thing over and over again for work you did once is the absurdity. And ridiculously greedy.
It seems to have been invented by laywers, for lawyers. Nobody else really benefits as much as they do. The whole entirety of society vs. a single profession of dubious morality.
>Somehow this idea that you keep getting paid for literally the same thing over and over again for work you did once is the absurdity
meanwhile, most tech is moving towards subscriptions?
Art is getting paid "non-greedily". People buy a song or art piece, and then people 10 years later buy a song or art piece. That's not one person paying twice for the same song, it's two people buying the same thing.
If people still value that art for that price later, I don't see how this is a "greedy" thing. is art magically supposed to turn open source CC0 after 5 years? Tech sure doesnt work like that.
You do other things to make money and continue to make art for it's own sake. If you get to the point that others want your art you get commissions. Just like most artists in the current system.
Alright, so then you're NOT going to pay artists for their labor?
In the current system, artists might work for many years on a single work, or work many years perfecting their craft before anyone wants to pay for their work. Copyright gives them a way to earn money in the future that compensates them for the work they did in the past. It incentivizes creativity. Don't get me wrong, I don't think copyright is perfect, but you really ought to think more about the system you're proposing, because it's not making much sense.
Unfortunately, it’s hard to explain these things to techies who only see the world in their one-sided startupy way. The fact that there’re starving creatives who have already been massively marginalized by the likes of Spotify of this world, means nothing to these tech workers who only see everything as numbers, or a “business model” to “validate”.
(full disclosure, I’m a techie who’s gradually woken up to the idea that the tech might just be the most abused way to exploit people)
I'm in games, where art and tech crossroad. I 1000% empathize for the fact that art exploits, abuses, and underpays even if they at times may be doing more work than a junior web dev.
It's a bit ironic, because a lot of tech offers partial compensation in stock. Something else that really doesn't happen in games unless you work for like, the 3-4 largest studios. So they should at least understand that your compensation is not all based on labor for time worked.
>You should be paid the accurate value of the labor.
that's gone out the door in the digital age. Compaies at this point have spent centuries trying to enfoce this model while witholding stuff like stock and royalties to take a part of what the company enjoys by protifting for decades off of a single (underpaid) piece of labor.
I don't exactly sympathize with a robot now trying to do the same. Pay your labor.
> How do you know what the value of the art will be before it's created?
I don't know. Anyone funding the work is accepting a risk.
> Should they not have been paid after 1991?
They definitely should get paid for their shows and live performances. The band itself can't be copied. Artists are extremely scarce.
Their art, however, is not. Once created, the scarcity of their recordings is artificial and fundamentally time limited anyway. Even if I were extremely tolerant of copyright, I'd argue for a term of only 5-10 years maximum with absolutely no possibility of extension.
In other words, even if we accept copyright as legitimate, they sure as hell shouldn't still be getting paid for some late 80s album. They've already been adequately compensated for those creations. If they want more, they should have to keep making new stuff so that they can benefit from new copyrights which will also expire after a short time.
Creators are not supposed to be able to strike gold once and then enjoy eternal royalties. Copyright must have short time frames or it's in breach of the social contract. The reality is we're doing creators a favor by pretending that it's hard to copy their stuff so they can make some money. We do this because they assured us that eventually all of it would belong to us: works would the public domain.
The copyright industry isn't keeping up their end of the bargain. They continuously pull the rug out from under us by extending copyright to the point we'll be long dead before our culture is returned to us. It's offensive and we should all stop pretending. They need reminding that public domain is the natural and default state of all intellectual work.
> How would you predict that value before its creation (or even after)?
I'd look at the artist's past work. If there is no past work, then I don't know.
> If you're saying that only the labor has value
I'm not saying that at all. Creations are valuable. Creators are valuable. The labor of creation is valuable.
Value is assigned to stuff by humans. Obviously humans value art. The price however is given by supply and demand. The fact is that supply of intellectual works approach infinity after they are created and therefore their prices approach zero. So it makes perfect sense to assign prices to the labor of creation but zero sense to assign a price to the product of creation. Copyright is an exercise in denying reality.
> and all labor is valued equally
I definitely did not say that. All labor is different. I value some creators a lot more than others. Some creators I don't value at all.
> that sounds sort of like marxism
I must apologize if I gave that impression. I hate marxism.
>In other words, even if we accept copyright as legitimate, they sure as hell shouldn't still be getting paid for some late 80s album.
Why not? The fact is that even if the album is free, there will be people paying spotify $10/month to listen to it on demand. How is it fair that Spotify can profit from it for decades to come because they offer convenience, over the artist who made the music 10 years earlier and now relinquishes their art not even a quarter into a typical career?
Copyright is abusrd now, but it's not a bad concept. I think the original copyright law of 14 + 14 worked well enough. Life expectancy increased so I'd increase it to 14 + 14 + 14 (or 10 years after the death of the original author, whatever comes first). You fund an artist for their typical career length (if they choose to extend twice) and once they are (near) retired the song is free to work off of. In the meantime you simply negotiate if you want to use their work.
The market at large will determine that. If people value cheap AI generated images more than talented human curated art then that's what it will be. If a market exists to buy unique pieces where an artist put brush to canvas and priced their work at $1000 instead of the cheap $10 poster that can be mass produced then thats what it will be. If no one wants to pay $1000 for your unique piece, then the market has spoken and your art is not worth that much. Like everything else, an equilibrium ill be reached. Good artists will be fine. The other 99% of self declared artists will fade away into obscurity.
None of that is what copyright protects. And it lessens the argument when you can argue that LLMs are essentially stealing a human art's work to be used to generate cheap images. Similar to how if you took commissioned art, printed out 1000 copies, and sold them for $1 a piece.
Copyright means that you need to at least pay that artist you stole from in some way, which the government enforces so artists don't stop creating.
First, you missed the "and". Do CliffNotes, Wikipedia, etc. substantially impact the market for the original work? For example CliffNotes does not - people who buy the CliffNotes version typically already have the original work as well (for example from coursework). And Wikipedia may well do more to interest people in the original work than to replace it.
Second, you ignored the "purely derivative" bit. You have to look at to what extent the use is derivative or transformative. See https://en.wikipedia.org/wiki/Transformative_use for a bit about that. (Note, this is a legal term defined by various precedents. OpenAI can't just argue, "Turning it into an LLM is a transform, so it is transformative!") Since CliffNotes is educational and Wikipedia is nonprofit, it is relatively easy for both to qualify as transformative.
As a result your response underscores the point that was made. There are a lot of shades of grey. You really can't just seize on a couple of phrases and key points, then jump straight to the answer. You have to understand how the courts will decide, and then accept that there is an actual judgment call whose outcome depends on the judge judging.
(I'm not a lawyer, but I have had excessive exposure to them in the past.)
The question was NOT whether it spreads information from the articles to people who wouldn't have paid for it. The question was whether it suppresses sales of the articles to people who otherwise might have paid for it.
That's a more complicated question of fact. Some people now read Wikipedia and won't buy the article. Some people encounter the reference on Wikipedia and decide to buy the article. Which happens more?
Publishers concluded that Wikipedia references are good for sales. And so jumped on the chance to cooperate with https://wikipedialibrary.wmflabs.org/. Which is therefore able to give free access to 90% of subscription only databases to you if you can prove that you're the kind of person who is likely to add citations to Wikipedia.
Legal questions are funny like that. You have to answer the question actually asked. If you merely answer another one that sounds similar to you, your answer is generally wrong.
I thought the question would be how it does this. If it can write NYT articles because it read them it has to arrive at the exact same words in the same sequence. Wikipedia has to copy and paste to achieve the same. So maybe the question actually asked does not apply.
> Questions of fair use are famously gray, and anyone who declares something as "entirely fair use", with no caveats, is nearly always wrong except for the must obvious cases, which the given example is most definitely not. A judge has wide latitude in determining fair use.
You're the one presenting unfounded claims with confidence here. There is well established case law about not being able to copyright facts. If you are actually fully paraphrasing a presentation of facts / ideas and not just altering a couple of words here and there, then there is a very strong case for non-infringement.
> You're the one presenting unfounded claims with confidence here.
No, I'm not. On the contrary, I'm really looking forward to this case because I believe it will be a great test of a bunch of concepts that are totally novel in the world of copyright law as it applies to generative AI. The only things I am presenting with confidence are:
1. That anyone who declares that something is unambiguously fair use (or, contrarily, unambiguously infringing) is likely wrong. There is simply too much latitude by judges, and there have certainly been cases where a ruling went one way, only to be overturned on appeal.
2. While I certainly have an opinion on how I think this case will be decided, I'm not presenting that with unwarranted confidence. Instead, I linked that great article on the 4 factors of fair use determination because it's clear to me lots of people are saying "fair use!" on one side or the other with no understanding of the factors judges must actually consider when making a determination.
You seem to be shifting the topic of this thread. The GP comment is about paraphrasing news articles while I don't see anything in the NYT lawsuit about paraphrasing. Rather, the NYT is concerned with exact reproduction or near exact reproduction. I too am very curious about the outcome of this case and wouldn't care bet either way on the outcome. I do have an opinion on what precedent would be better for our society but that doesn't mean I think that outcome is more likely.
However, none of that matters in this particular thread. There are well established precedents about paraphrasing news articles and they do not support the claim you made
The "unfounded claims" were backed up by a link to Stanford on fair use and copyright. That's the opposite of being unfounded.
Remember. The NY Times does not have a record of filing frivolous lawsuits. Particularly not against companies with deep pockets. So it is almost certainly true that a lawyer who knows the law better than you thinks that this has a real chance. So you should be looking for flaws in trivial defenses that you can think up, rather than assuming that you know best.
For example take your copyright facts defense. That would be great if the NY Times was a phone book. They aren't, in addition to facts they offer analysis, editorial positions, and so on. For example I just asked ChatGPT, "In 2016, did the New York Times generally support or oppose President Trump?" I got back an answer talking about various kinds of concerns that the New York Times had, including an editorial titled, "Why Donald Trump Should Not Be President". The copy that ChatGPT needed to have to do that has a lot more than just facts in it.
Now if you paraphrased the NY Times like ChatGPT did when it answered me, you'd have a perfect fair use defense. But you aren't doing it for money, you didn't make a copy of all the NY Times, you aren't destroying the market for the NY Times, and you're legally able to own copyright in your transformed work. OpenAI is doing it for money, did copy all of the NY Times, is seriously impacting the market for NY Times articles, and ChatGPT generated text does not get a copyright.
Fair use is filled with shades of grey. Even if ChatGPT appears to do the same thing that you do, it is far less clear that OpenAI will enjoy the same level of fair use defense.
The Stanford link is just generic information about the fair use tests and does nothing to backup the assertion.
> They aren't, in addition to facts they offer analysis, editorial positions, and so on.
Those opinions and ideas are also not copyrightable. Only expressions of them are copyrightable, which is why paraphrasing facts, ideas and opinions is not a violation of copyright.
> Fair use is filled with shades of grey.
Yes, but not all those shade are equal. There is a long history of litigation showing that paraphrasing news articles is fine.
This is the weakest part of the case(s) against OpenAI. "Derivative work" is a legal term of art meaning a direct adaptation, like writing a screenplay of a book or translating a book into another language.
NYT has a stronger case than Sarah Silverman here because they can show actual 'memorized' text rather than just summarization, but given that those memorizations are a) an unintended failure mode of the training process, and b) from an older version of the model that has been updated to no longer regurgitate memorized text, it's not really clear how in current form GPT could possibly be considered a derivative work.
A question is whether the new model still intrinsically embeds the source text, but this is later filtered in the output, or if it no longer embeds the text at all.
I would think an existing model could bootstrap a copyright free training corpus by completely rewriting/paraphrasing copyrighted material with semantic fidelity for training of the next model to completely eliminate memorization of copyrighted works. That might pose an interesting obstacle to copyright challenges, bootstrapping your way into a clean room. Although, tweaking the architecture to either eliminate memorization, or eliminate high fidelity reproduction of verbatim training data seems far more expedient and less costly.
"Transformative" seems to fit a lot more that "Derivative".
On the other hand, it's understandable why NYT is worried. OpenAI itself says that occupations like: Writers and Authors, Web and Digital Interface Designers, News Analysts, Reporters, and Journalists, Proofreaders and Copy Markers
are "90-100% exposed" to what OpenAI is building.
Yup. News has been rampant with speculation, hearsay, and propaganda all my life。 Content mills and astroturfers already bury the truth or relevant stories with noise.
I don't buy into all these "dangers". The advent of cars did not decrease the amount of drivers and introduced various new jobs, that were not available for a lot of people. And the rise of computers, did not make the workforce smaller but instead opened many more opportunities for a lot of people.
I think focusing on lost jobs is the wrong angle to take this in. It's in how the content being used is being compensated. OpenAi isn't paying writers to train their engine.
I don't care that the car replaced the horse carriage because it didn't need to compensate horses nor handlers to do so. AI being the newest iteration of scraping data from artists, writers, etc. to profit millions off of is directly using the "horse handler's" work. If these LLM's threw NYT a royalty to use their articles as training material, there wouldn't be a lawsuit.
I personally appreciate the semi truck sized loophole that is satire. One can include an entire copy written work within one's own work as long as the treatment of that other copy written work is parody / satire. This is a provision of US copyright law put in place to protect political satire, which can be anything, because politics is everything.
I would say it is arguable that is fair use, but the whole thing about fair use is that it is a defense, not a type of license or something you can preemptively apply. So whether or not it will be protected under fair use is actually not determined yet. In fact I would say that’s the entire debate here, right?
I have worked on many documentaries and any time we said “fair use” internally what we were implicitly saying is “nobody will come after us because they know that we are probably safe under fair use if this escalated.“ But again, we could never preemptively apply it. We were just anticipating potential conflict and gauging how likely it was to occur.
Why do you say that? Commercial vs noncommercial use is a primary factor in the “purpose” prong of the fair use balancing test and a significant one in the “market effects” prong.
That a use is noncommercial is often a deciding factor in the success of a fair use defense. GP is overstating it though, since it’s still one of many factors.
Because anyone that is familiar with fair use knows that the purpose prong and the commerciality aspect of it is not one of the more important prongs of the fair use analysis, whereas transformation is. Transformation adjusts what is a purpose that falls under fair use. Did you read Warhol??
Yes. Warhol is an example where the commercial nature of the secondary use was the deciding factor in its failure to pass the purpose prong.
> In sum, if an original work and secondary use share the same or highly similar purposes, and the secondary use is commercial, the first fair use factor is likely to weigh against fair use, absent some other justification for copying.
(P4). It’s very likely that a noncommercial secondary use would have passed under the reasoning in Warhol. I don’t understand the point you’re trying to make.
Read what you quoted - the commerciality of the use comes after whether or not the use was transformational. That's the entire of point Warhol - when the use is not transformational, there is very little space for a commercial fair use.
Whether or not the use is commercial is certainly one of the considerations, but it's not the most significant one generally. There certainly can be specific cases where it's very significant, of course.
But what I was arguing was that a use is not "fair use" merely because it's noncommercial in nature. I cannot make copies of movies and give them away on the street for free and successfully claim "fair use".
He doesn't actually make very heavy use of the satire plank of fair use. He credits the original artists. From his own website
"Does Al get permission to do his parodies?
Al does get permission from the original writers of the songs that he parodies. While the law supports his ability to parody without permission, he feels it’s important to maintain the relationships that he’s built with artists and writers over the years. Plus, Al wants to make sure that he gets his songwriter credit (as writer of new lyrics) as well as his rightful share of the royalties."
The fact that he could rely on fair use is separate from whether he as an artist does rely on fair use.
I'm pretty sure Weird Al is actually using compulsory licensing in music and just paying the required royalties to the songwriters. Anyone can cover any published song, you just have to pay the royalties when you do.
Always great to see people point out Weird Al, cause he's the shining beacon of an example of what OpenAI et al. should be doing. He explicitly gets permission from the original authors before doing any of his parodies, and he's even been turned down a few times as well, famously Prince rejected him a bunch of times and he subsequently has never made a Prince parody.
Not only does he get permission from the original authors, he also pays royalties to the authors despite legally not having to do so.
If that were true, I could take a band that I hate, copy all of their music note-for-note, then release an exact copy on the market and undercut them by selling their entire discography for $0.01
Fair Use requires one of several enumerated activities, including satire, education, journalism. You can’t just copy content and hope that it passes Fair Use.
Hire a lawyer if you are unsure. But at least read the Wikipedia article on the subject if you are going to talk about it.
He's talking about citing and quoting NYTimes articles, not republishing them verbatim. That said, it's very different if you're a publication that sometimes cites reporting from other publications vs. a website exclusively dedicated to indexing and summarizing NYTimes articles.
I couldn't get gpt to quote an actual nyt article no matter how hard I tried...it just hallucinated in the general style of a news article.
Presumably, if it can remember at least a paragraph or two of each article, then surely the same would be true of any text it ingested and the model size would approach the dataset size (probably actually much larger). I don't believe this is the case at all, even searching around, I've not found any good recent examples of it regurgitating copyrighted text verbatim.
It's cool to hate AI stuff if you're a creative atm. But gotta love those generative/algorithm based PS brushes, that's still real art!
"Indeed, the opening paragraph of "A Game of Thrones" by George R.R. Martin, with the chapter titled "Bran," starts as follows:
"The morning had dawned clear and cold, with a crispness that hinted"
And then it cuts off, whether that's because OAI now have an oh shit filter or just the model had access to the first page or publicly available articles quoting the first line, I'm not sure.
I tried other chapters and random sections and it could get a sentence or two right but then hallucinated; what's more likely NYT and GRRM? That your works are being reproduced verbatim? Or that Facebook, YouTube descriptions, fan tumblrs and hell, the publicly available and multiple GoT related wikis that include a variety of passages from the books were used as training data?
I don't think it's necessarily true that model size would need to be larger than dataset size. It's theoretically possible that the model encoding achieves significantly better compression than DEFLATE or GZIP or whatever compression algorithm is used to store the dataset itself.
I think what wouldn't be covered is reproducing substantial portions of an article, especially if it's done without attribution. Tier 2 publications that fully reprint NYT or AP/Reuters articles are usually doing this via a paid News Service or Content License. See: https://nytlicensing.com/content/new-york-times-news-service...
Correct. But those 2nd tier sources don't have the NYT copy verbatim. Do you really think the US NFL, as an example, would let OpenAI use all of its recorded games as a way to train some new GenAI game framework to build better American Football games? No. All that material is copyright. Public media is going to move to a very awkward era of ownership and licensing because all of these large companies looking to make a buck off public data sets are doing very little to make the economic model less one sided.
I hope the NYT prevails here, personally. Models will (and are) currently tainted by data they should not contain and for longer term privacy concerns this needs to be addressed early and have significant consequences or we're headed towards a world where this type of technology will make our ad-targeted world seem like a much more manageable past.
> To keep things simple, let's say I never regurgitate chunks of verbatim NYT articles, maybe quite short snippets.
You just described Google. When you think about it, it's surprising that Google is legal. However, it is well established that what Google does is perfectly legal. Remember that internally Google keeps and uses complete verbatim copies of every web page they index.
Yes, Google offers a link to the source. If OpenAI did the same, even if only 0.1% of people clicked on the links and NYTimes hardly got any revenue from it, would that make it legal in your eyes? What if they implemented a system that detected when it was outputting a verbatim copy of something and simply paraphrased it? NYTimes clearly doesn't have copyright on paraphrased versions of their articles. I think it would be pretty silly if the government forced them to do that as it wouldn't make any practical difference to anyone.
> However, it is well established that what Google does is perfectly legal.
Google has a wide range of products and shakedowns. Not all of them are "perfectly" legal: Google is being challenged in court over some of their shakedowns and products practices.
I am clearly talking about the web search engine in the context of copyright. Other products or legal concerns like antitrust are completely irrelevant here.
Any publisher can opt out of google. Publisher also have substantial control over titles and snippets shown in google, whether an article appears in google news, etc
Paraphrasing is also known as cloning and is often a copyright violation
Copyright law doesn't mention opt outs or search engine snippet controls. It's not clear to me that robots.txt is the singular thing that makes Google legal.
In US copyright law facts cannot be copyrighted, so copyright on factual content like newspaper articles is limited. Simply replacing a few words wouldn't work, but I am certain that GPT-4 is capable of paraphrasing factual content at a level that would not be considered infringement if a human did it.
>Copyright law doesn't mention opt outs or search engine snippet controls. It's not clear to me that robots.txt is the singular thing that makes Google legal.
Genuinely - what are you talking about besides your own assumptions? you just assume everything google does is legal and therefore any one else doing anything arguably similar must also be legal? Without regard for factual details that do matter to copyright law? Such as license?? Your own description of copyright law here is very stunted - you can't paraphrase articles of the NYTimes and call it a fair use. You can report on what the NYtimes reports on... because that's what news is.
If I make a website that scrapes NYT and passes it back and forth through a machine translator, say, English -> Spanish -> English, then the content will be slightly modified. Is this legal to make money off of?
Seems like the legal answer is unclear but, like Napster, such a system seems like it would lose in court.
It would be unlikely to be something you'd find paying customers for, though? I suppose if you charged a small percentage of what NYT charges people might be willing to consider it, but you'd have some costs for hosting etc., so I am skeptical about its viability as a business model...
I'd serve fake news en-masse to low IQ people who click things to feel good about their own views. I'd also build a handful of websites (ideally as many as I can personally manage) to flood the Internet with fake news clickbait.
One site clones fox news. One clones news max. And so on, cloning many news sites, sports sites, any news site. Automated, massive scale content farming. Think of the websites recommended by Taboola but, realistically, a whole lot worse.
That’s not the only reason. Google search is also transformative and non competitive with the underlying publications. And that is why the opt out is important. If you feel google competes with your site you don’t have to sue Google: just tell them to to away
Transformative yes, so is ChatGPT. Much more so actually. Non-competitive is debatable. Especially with the instant answers Google has in addition to regular snippets which can also obviate the need to visit a site. I have a hard time seeing ChatGPT as competing with newspapers more than Google Search does.
Nobody is seriously going to ChatGPT and trying to trick it into regurgitating old NYT articles as an alternative to paying for access to NYT's archives. Meanwhile, newspapers went as far as getting the laws changed in several countries because they felt Google was competing with them too much and didn't like the fact that it was legal.
Can they? Here's reference to a legal fight where Google scraped song lyrics from a lyrics website, and presented the lyrics verbatim directly to users (bypassing the original site and the ads that allowed that site to operate)
You took that quote out of context and missed the broader point in the process. The snippets provided in regular search results cannot generally replace the substance of the full articles they link to, while that's the whole point of GP's hypothetical website—it simply doesn't reproduce large chunks of text verbatim, presumably to avoid copyright infringement claims in the hypothetical's frame, and in GP's rhetorical frame to present an analogy with the information-laundering powers of LLMs that their creators claim make their exploitation of unlicensed training data fair use.
The whole point of a search engine (as we've classically known them) is to index the web and respond to queries with a list of links that you will inspect and click through on. The whole point of an LLM chatbot tool is to eliminate those inspecting and clicking-through steps, becoming a one-stop shop for content whose substance was created by someone else. That's also the whole point of GP's hypothetical, which is why it works as an analogy.
---
There are substantially better arguments for search engines being legitimate fair use. Consider, for example, transformation. AI defenders will argue that these systems are transformative because they reshuffle elements of their input in their output, but that's clearly a much weaker form of transformation than one in which the transformed work has an entirely different nature and purpose, i.e. search engines vs. the results they return. Ultimately these technicality-based "nuh uh" arguments aren't going to save the practice of training AI on unlicensed data, because they are incompatible with the spirit of copyright law even if the novel nature of these technologies means the letter of said law can't quite nail them down yet.
If these arguments do succeed, it will be because the judicial/regulatory environment in which they were applied has been corrupted by capital.
A search engine takes an input string, a corpus of text, and returns a series of text that best comes next after the input string.
An LLM takes an input string a corpus of text, and returns a series of text that best comes after the next input string.
To get a paragraph of output, you run the search over and over again
Both the search and LLM reshuffle the inputs to the outputs.
If I'm describing the purpose of the LLM, it's got a wide number of usages. "Making my resume look more professional" or "be a crud api" or "reformat my ask into a api call to X service" or "give me a timeline of events surrounding Y with source links"
An 18 wheeler travels on wheels. A shopping cart travels on wheels. A shopping cart does not require a license to operate, therefore an 18 wheeler does not require a license to operate. A shopping cart can be operated inside a grocery store, therefore an 18 wheeler can be operated inside a grocery store. A child can operate a shopping cart, therefore a child can operate an 18 wheeler.
If I'm describing the purpose of an 18 wheeler, it's got a wide number of usages. "Carry my chicken" or "carry my lettuce" or "carry my Cheetos". Or, simply, "carry my groceries".
> And how did the training data contribute to the content in any meaningful way? Inspiration isn't substance.
If training data is so unimportant, why not simply not use it and avoid the controversy? At the very least that would certainly fix the issue where the model demonstrates how "inspired" it is by NYT articles by reproducing them verbatim.
The reason why Google keeping entire digital copies of other people's copyrighted works is legal is because copyright is all about distribution rights. Any person can possess the entire works of Disney (without paying for them), for example and as long as they do not distribute those works they're 100% in the clear.
Possession is not a crime when it comes to copyright. It's not like physical things (e.g. drugs or guns) at all. This is why comparing copyright violations to theft is silly.
ChatGPT can absolutely keep verbatim copies of the entire works of basically anything and not run afoul of the law. When it regurgitates a small part of an article that's covered by fair use in theory but the truth is that fair use can only be determined by a judge in a court of law when someone is sued. It cannot be determined with any sort of certainty ahead of time. It's a legal defense, nothing more.
Summarizing content has been legal forever as well (see the other posts here talking about Cliff Notes and some similar products). That's not even fair use that's just like, people's opinions, man (legally speaking).
I don't think the NYT will get what they want out of this at all.
I agree with your IANAL take, but what about a situation with an extra level of indirection? So the service never reads actual NYT articles, but only reads blog/forum posts about NYT articles, and derives what is in the article from conversations about the article by people who have read it. Is that legal now?
This analogy fails to capture the transformative nature of these models. Hosting a derivative work that is also a news article is not transformative. Hosting a next word completer is very different than a news article and can't be used as a substitute.
> If you forget about the LLM aspect, and simply build a product out of (legally) scraped NYT articles, is that fair use?
That's not a good question.
If I look out of my window and see my neighbor go to the shop, that's fine. If I use cameras and track everybody I see on the street and put them in a database, then that's problematic and illegal in many places.
Logic does not necessarily apply when scaling is involved.
It depends. Google built a product out of scraping content (Google Search).
But what I'm saying is that answering the question does not allow you to deduce anything about your rights; that's what I mean by "not a good question".
It can allow you to deduce something, depending on the answer.
If we want to establish whether scenario A is fair use or not, and we all agree that A is "worse" (regarding fair use status) than some other scenario B, then if we also agree that B is not fair use, A by definition isn't either. The opposite is not true, of course: B being fair use does not imply that A has to be as well.
I find that kind of upper/lower bound logic can be pretty useful and I think it's what the parent comment was trying to do.
On a related note, that same logic is why I think Godwin's law can be a bit misapplied now and then. Sometimes bringing up nazis/Hitler can be useful to establish some ground truth in a debate (instead of just a way to imply your opponent is actually a bad person, or, possibly, an actual nazi themselves). E.g. a conversation on the morality of violence is vastly different depending on whether you agree that violence against nazis is ok or not.
I think it can still provide value if the actual scenario at hand is so complex and fraught that conversations about it end up mostly fruitless (as I think is the case here). At least it can provide you with some mental handholds and supports for where to start reasoning about the problem, which hopefully helps in finding some small agreements, or at the very least, mutual understanding of each other's positions.
The general answer is no. Fair use is a special carve out legally that has to be determined individually. If your product is something that regurgitates NYT articles while stripping NYT of their source of revenue, that’s got fair odds to not qualify as fair use.
> If I use cameras and track everybody I see on the street and put them in a database, then that's problematic and illegal in many places.
Afaik not illegal in the US. You put a camera on your own private property (window), use it to record what’s happening in a public space (the street outside), and then store that data in a database (that other people can presumably access). Unless I am missing something, this scenario is perfectly legal in the US.
Inbefore I get hit with “not every country is like that at all,” NYT is based in the US and the lawsuit is filed in the US. So how a bunch of other countries deal with similar issues shouldn’t really have as much bearing on this specific case.
> Implications: The Ninth Circuit's declaration that selectively banning potential competitors from accessing and using data that is publicly available can be considered unfair competition under California law may have large implication for antitrust law. [citation needed]
> Other countries with laws to prevent monopolistic practices or anti-trust laws may also see similar disputes and prospectively judgements hailing commercial use of publicly accessible information. While there is global precedence by virtue of large companies such as Thomson Reuters, Bloomberg or Google [or LexisNexis or Westlaw] effectively using web-scraping or crawling to aggregate information from disparate sources across the web, fundamentally the judgement by Ninth Circuit fortifies the lack of enforceability of browse-wrap agreements over conduct of trade using publicly available information.
IANAL but aren't the key terms there "selectively banning" and "publicly available"?
NYT articles are largely behind a paywall for everyone. That means they are not publicly available, and a competitor who was blocked from accessing or reproducing that content without a license would not be "selectively banned"
I believe that part of the law suit contends that the content wasn’t able to be scraped “legally” as you put it. Instead they show that ChatGPT will regurgitate verbatim excerpts from articles that are behind the paywall.
If there is payment then usually there is an agreement. An agreement can limit fair use. Can the NYT, via an agreement, e.g., "Terms of Use", limit what the subscriber does with the articles. There is not much precedent that suggests otherwise.
Consider the analogy from libraries that want to do data mining.
"Unfortunately, in licenses for digital scholarly content the majority of content acquired by research libraries publishers often include terms that prohibit certain uses that would otherwise be allowable under the Copyright Act. For instance, licenses may require libraries or individual researchers to negotiate for otherwise lawful activities, such as text and data mining, and to pay exorbitant fees on top of the cost of the content itself. While new regulations allow researchers to circumvent technological protection measures to access copyrighted materials, licenses for that content may include terms that explicitly prohibit this circumvention. In many cases, these activities might actually increase the value of published material; for instance, if a data-mining project yields new knowledge about a topic covered in a journal, it may very well spark new interest in that journals content. Libraries and publishers have often assumed that license terms that restrict copyright exceptions are enforceable under state contract law. There is, however, surprisingly little case law on this point."
Putting some string in a robots.txt to try to stop data collection is an amusing "solution". Should copyright owners have "Terms of Use" that limit usage for commercial "AI" purposes.
NYT's perspective is going to look so stupid in future when we put LLMs into mechanical bodies with the ability to interact with the physical world, and to learn/update their weights live. It would make it completely illegal for such a robot to read/watch/listen to any copyrighted material; no watching TV, no reading library books, no browsing the internet, because in doing so it could memorise some copyrighted content.
I disagree. The verbatim part is the problem. You’re drawing a comparison to how humans operate except we’re not allowed to operate like that.
While harder to do as a human, if memorised a copyrighted book and then did a live reading on TV, or produced replicas from memory and sold them (the most comparable example), I’d be sued.
Humans produce derivative work all the time, and it’s fine for LLM’s to do that, but you can’t do it verbatim.
>or produced replicas from memory and sold them (the most comparable example), I’d be sued.
This is not the most comparable example, because it's not what ChatGPT is doing. The most comparable example is if you were hired as a contractor and the employer asked you to write verbatim some copyright content you'd memorised. If the employer then published it, they'd be the one liable, not you.
>Humans produce derivative work all the time, and it’s fine for LLM’s to do that, but you can’t do it verbatim.
Nobody's suggesting preventing humans from consuming any copyrighted content just because in future they might recite some of it verbatim, but that's what NYT want for LLMs.
> The most comparable example is if you were hired as a contractor and the employer asked you to write verbatim some copyright content you'd memorised. If the employer then published it, they'd be the one liable, not you.
No, you'd both be liable. You are not allowed to create copies of a copyrighted work, even from memory, for any commercial purpose. Making it public or not is irrelevant.
This is more obvious with spftware: if I copy a version of AutoCAD that my previous employer bought and sell it to another company, or even just use it for my current employer without showing it to anyone else, I am violating the copyright on that software, and I am liable. Even though obviously no "publishing" happened.
Similarly, if you hire a decorator to paint Mickey Mouse on the inside walls of your private kindergarten, the decorator is violating Disney's copyright just as much as you are, even if neither of you has made that public.
Your previous employer never bought AutoCAD, they licenced its use, paying a subscription. When you start working for them that licence was no longer available to you. So you would be unable to subsequently use it.
Unable legally, but I may find illegal ways. And the reason it is illegal to copy is copyright at the end. The license is only (legally) required because of copyright.
Then we should be focused on policing the usage of the model, not the training of it.
That's the point at which infringement occurs in your example. It's not the memorizing that's the infringement, it's the reproduction from your memory.
We shouldn't be regulating your hippocampus encoding the book, but your reproducing the book from that encoding.
Similarly, we shouldn't be regulating the encoding of material into the NN, but the NN spitting back out the material.
Are those LLMs independant citizens we are going to give rights to? Then I'm fine with that.
Are they all owned by one mega-corporation, which is going to do as capitalism does, and use them to squeeze money out of all of us? Then I'm happy to ban them.
"Let's ban something capable of diagnosing medical conditions and letting coma patients to communicate with an EEG because it learned the relationships between words from a giant data set of scraped data and is owned by a company" is a pretty callous take IMO.
The opportunity cost of holding this technology back is going to literally be millions of people's lives given current trends in its emerging applications.
Memorising isn't the issue. It's providing it back verbatim and/or cutting access to the source.
You'd get the same problem with someone with a photographic memory who a group of people would turn to recite them the news instead of buying the newspaper.
As of now public performance of copyrighted material is infringement.
That's not the case, as they aren't trying to get a ruling on the forced reproduction by prompt as infringement, but rather to get a ruling that training is infringement.
I fully agree with the perspective that infringement in usage needs to be limited even if I strongly disagree that training is infringement.
I read about this in the Times today (and am surprised that it wasn't on HN already).
My guess is that the court will likely find in the Times favor, because the legal system won't be able to understand how training works and because people are "scared" of AI. To me, reading a book, putting it in some storage system, and then recalling it to form future thoughts is fair use. It's what we all do all the time, and I think that's exactly what training is. I might say something like "I, for one, welcome our new LLM overlords". Am I infringing the copyright of The Simpsons? No.
I am guessing some technicality like a terms-of-use violation of the website (avoidable if you go to the library and type in back issues of the Times), or storing the text between training sessions is what will do OpenAI in here. The legal system has never been particularly comfortable with how computers work; for example, the only reason EULAs work is because you "copy" software when your OS reads the program off of disk into memory (and from memory into cache, and from cache into registers). That would be copyright infringement according to courts, so you have to agree to a license to get that permission.
I think the precedent on copyright law is way off base, granting too much power to authors and too little to user. But because it's so favorable towards "rightsholders", I expect the Times to prevail here.
My hard drive can - bit for bit - recall video files. If I serve them to other people on the internet without permission of the copyright holder, that’s called piracy.
Yeah, but the LLMs can't. They aren't big enough to contain every byte of every NYT article, even with the best-known compression algorithms. Rather, they pick up and remember the same patterns that humans do when they write. Authors of the articles also did that, and so the two algorithms (human writer, LLM inference) end up with the same result. (That doesn't preclude large chunks of text that are actually remembered, though. We humans have large chunks of verbatim text floating around in our brains. Passwords, phone numbers, "I pledge allegiance to the flag...", etc.)
Anyway, like I said, I don't think OpenAI will win this. Someone will produce one verbatim article and the court will make OpenAI pay a bunch of money as though every article could be reproduced verbatim, and AI in the US will be set back that many billion dollars. It probably doesn't matter in the long run; it preserves the status quo for as long as the judge is judging and the newspaper exec is newspaper exec-ing. That's all they need. The next generation will have to figure out how to deal with AI-induced job loss... and climate change. Have fun, next generation!
In general, if you perform copyrighted works you are doing copyright infringement. There are certain exceptions (personal use, education, very small fragments with proper attribution, maybe a few others) but whether you are reading it aloud from a book or performing it from memory makes no difference.
So, if you setup a service like ChatGPT but powered by humans responding real time to queries, and these humans would occasionally reproduce large chunks of NYT articles, they and the service itself would be liable for copyright infringement. Even if they were all reproducing these from memory.
Now, this is somewhat different from the discussion of whether training the model on the copyrighted data, even if it had effective protections from returning copies of it, constitutes copyright infringement in itself. I believe this is a somewhat novel legal question and I can think of no direct corollaries.
I certainly don't think we can just handwave and say "at some level, when a human reads a copyrighted work, they are doing the same thing", because we really don't know if that is true. Artifical neural networks certainly have no direct similarity with the neural networks in the brain as far as we can tell. And, even if they did, there is no reason to give a machine the same rights that a human has - certainly not until that machine can prove sentience.
It's extremely speculative to claim that LLM models are basically doing what humans do. There is very clearly something that isn't right about that because in order for a human to learn to speak and converse and they don't need to imbibe the entire corpus of all written text in human history - which is basically what we're doing with these LLMs. What we're giving them is vast amounts of data which is totally unlike how humans work. There's very clearly some gap here between what a LLM is doing and what a human is doing. So you can't use that as a basis to justify why it's ok for OpenAI to operate like this.
To put it another way, let's say I turn the dial all the way the other way, I train the worlds crappest LLM on NYT material, it massively massively overfits and all it will ever return is verbatim snippets of the NYT. Is that copyright infringement?
The core part of the argument here is actually just that OpenAI doesn't want to adhere to what the current standard is for using copyrighted material, if you want to use it and create something new with it you need to license the material. Since OpenAI's LLM isn't actually like a human it needs to license such a vast dataset that it would be uneconomical to run the business without stealing all the content.
Isn't copyright tethered somehow to a notion of "expression"? That is, the same ideas and facts expressed differently are a different work?
Sure, when something is clearly derived, or just expressed in a new medium, then I'm sure it's still covered. But if it goes through an LLM and the result bears little resemblance, how can that still fall under copyright?
As you said AI can rewrite articles, obtaining a clean cut separation between ideas and expression. Keep the ideas, write a new text. And if you got multiple sources, the more sources you use the better, it would make the output be even more different. This approach could also check consistency and bias between sources.
920 comments
[ 3.7 ms ] story [ 176 ms ] thread"The suit seeks nothing less than the erasure of both any GPT instances that the parties have trained using material from the Times, as well as the destruction of the datasets that were used for the training. It also asks for a permanent injunction to prevent similar conduct in the future. The Times also wants money, lots and lots of money: "statutory damages, compensatory damages, restitution, disgorgement, and any other relief that may be permitted by law or equity.""
If we see court judgements start to go copyright owners way, we will also see a scramble from AI companies to buy the few publishers with enough data to be worth buying, and to create works for hire to replace the rest.
In the long run a copyright ruling like that will be a boon for OpenAI and all other players with deep enough pockets to do so, and massively harm everyone else who will suddenly find it far harder to build models legally.
[0] https://www.nytimes.com/2023/12/22/technology/apple-ai-news-...
[1] https://www.theverge.com/2023/12/22/24012730/apple-ai-models...
( https://en.wikipedia.org/wiki/HiQ_Labs_v._LinkedIn )
robots.txt on nytimes.com now disallows indexing by GPTBot, so there's an argument against automated information acquisition starting from some moment, but before some moment they weren't explicitly against that.
I do think that’s the case for some things but especially for new things that doesn’t seem like a common sense understanding of the world.
If you don't want people to get at your land, setting up even a small fence creates an explicit indication of limitations. Just like the record in robots.txt I mentioned earlier.
New York Times also doesn't limit article text content if you just request HTML, which is typical for automated cases. But they impose th limits imposed on users viewing the pages in browser with Javascript, CSS and everything else. So they clearly:
1. Have a way to determine the user's eligibility for reading the full article on server side.
2. Don't limit the content for typical automated cases on server side.
3. Have a way to track the activity of not logged in users, determining the eligibility for access. So it's reasonable to assume that they had records of repeated access from the same origin, but didn't impose any limitations before some time.
So there are enough reasons to think that robots are welcome to read the articles fully. I'm not talking about copyright violations here, only about the ability to receive the data.
The trajectory and value to society of OpenAI vs NYtimes could not be greater. They have won no favors in the court of public opinion with their frequent misinformation. It's all just a big waste of time, the last of the old guard flailing against the march of progress.
And even hypothetially if they managed to get OpenAI to delete ChatGPT they'd be hated forever.
You mean GPT here, right?
The NYT may produce misinformation but it aims not to, and its staff of human writers are limited in the quantity that they can produce. They also publish corrections.
GPT enables anyone who can pay to generate a virtually unlimited volume of misinformation, launder it into 'articles' with fake bylines and saturate the internet with garbage.
I think we need to focus on the damage done.
In that case the bigger danger is Open source LLM's. OpenAI at least monitors the use of their endpoints for obvious harm.
Except when it affects their bottom line of course, they publicly lied on how meta tags work during the lawsuits against Google to get more money (like most newspapers did). And I have no doubt that they will extensively lie once again on how LLM really work.
Train a model on NYT text that outputs a summary of facts that it learned: OMG literally murder.
Also remember copyright laws was not there in the first place.
If you can't copyright AI-generated pieces, then why would fair use apply to LLMs?
Is it? Can you quote relevant legislation or case law?
I read a NYT article and publish an exact copy of that article on my website: copyright infringement.
Train a model on NYT text and it outputs an exact copy of that text: also copyright infringement.
Probably not until they pay him a hefty copyright fee.
Determining whether a work violates a copyright requires holistic consideration of the similarity of the work to the copyrighted material, the purpose of the work, and the work’s impact on the copyright holder.
There is not an algorithm for this, cases are decided on by people.
There are algorithms that could detect obvious violations of copyright, such as the one you suggest which looks for exact matches to copyrighted material. However, there are many potential outputs, or patterns of output, which would be copyright violation and would not be caught by this trivial test.
What does that mean?
Look up "substantial non-infringing use" and this little court case:
https://en.wikipedia.org/wiki/Sony_Corp._of_America_v._Unive....
Now spend a few million on lawyers and roll your dice.
NYT won't mind if you use their content to train LLMs - as long as they get a commission. Reddit will shut down their free API and make you pay to get training content. Discord is going to be selling content for AI training too - if they haven't already done so. Twitter is doing it.
They didn't care before because LLMs were just experiments. Now we're talking trillions of dollars of value.
Can you make the argument this was their fault for not having forward vision/being asleep at the wheel and "accidentally, in hindsight" letting OpenAI/others have free, open, unlimited access to their content?
They are not giving it out "for free", in fact they're being paid by their employer to write these articles. Moreover, the writers themselves stand noth' to gain from their past writings financially as they don't belong to the ownership structure of the business.
This is a dumb argument. We're not just talking about ancient articles. We're talking about new content, including content that is yet to be written.
ChatGPT isn't competing with NYT on a core competency. No one uses LLMs for original news reporting. They're obviously incapable of doing that, by virtue of not being there on the scene or able to independently research a topic, maintain relationships with sources, etc. What ChatGPT can do is quote/reproduce some parts of past articles, and reason from them. Or at least produce new text that's somewhat related to the old text.
The threat to NYT is this: ChatGPT is much better bullshitter than they are, so it reduces NYT to its core competency: providing original information. Which is all it should be doing in the first place. But instead, NYT wants to not only keep the bullshitting part of its revenue, but also take a cut or destroy the much greater and much more useful part of where this all feeds a general-purpose language model.
This is a badly-formulated conjecture, or worse, ultimately selective reading of "social credit" which only purpose is serving your argument; it has nothing to do with economics. I'm sorry, but I'm not convinced.
OpenSource developers did that ;)
Earnestly I found ";)" deeply troublesome.
You're fighting a scarecrow that doesn't exist...
Can you say the same for user created content on Reddit, Twitter, or Facebook? A user agreement that nobody reads doesn't have anything like the same legal basis as a signed contract. Not to mention that a large percentage of users are not adults.
This tired 'fair use' excuses from AI bros whilst the GPT has reproduced the article text verbatim, word for word and it being monetized without the permission from the copyright holder and source (NYT) is an obvious copyright violation 101. Full stop.
Again, just like Getty v. Stability, this copyright lawsuit will end in a licensing deal. Apple played it smart with their choice with licensing deals to train their GPT [0]. But this time, OpenAI knew they could get a license to train on NYT articles but chose not to.
[0] https://9to5mac.com/2023/12/22/apple-wants-to-train-its-ai-w...
Stability took the easy way out because they didn't have billions of dollars to play around with and Microsoft to back them. Let's see what OpenAI does but calling everyone who disagrees with your naive interpretation of fair use "AI bros" is doing everyone a disservice.
What (or whom) do you consider to be an "AI bro?"
This sort of ad hominem generalization usually accompanies a weak argument.
And even if they did it will be fine because those sources allow for it.
The point is that OpenAI never asked NYT for permission to use their data.
Fair use has nothing to do with reproducibility. LLMs are more clearly fair use than a search engine cache and those court cases are long settled. There's no world in which OpenAI doesn't win this entire thing.
Why do you think the architecture is important? If I have a computer program and it outputs the an entire copyrighted poem then the answer to "is this copyright violation" SHOULD NOT depends on the architecture of the program.
It gets harder to stand behind a blanket claim that LLMs or any AI we’ve got falls under fair use when they keep repeatedly reproducing complete and identifiable individual works and clearly violating copyright laws in specific instances. The models might be remixing and/or transformative most of the time, but we have proof that they don’t do that every time nor all the time… yet. Maybe the lawsuits will be the impetus we need to fix the AIs so they don’t reproduce specific works, and thus make the fair use claim solid and actually defensible?
https://www.youtube.com/watch?v=eUHBPuHS-7s (the original is flash and has thus been consigned to the memory hole, so we are left with this poor-quality conversion)
36": 'however, the press as you know it has ceased to exist'
40": '20th-century news organizations are an afterthought; a lonely remnant of a not-too-distant past'
2'11": 'also in 2002, google launches google news, a news portal. news organizations cry foul. google news is edited entirely by computers'
5'13": 'the news wars of 2010 are notable for the fact that no actual news organizations take part. googlezon finally checkmates microsoft with a feature the software giant cannot match: using a new algorithm, googlezon's computers construct new stories, dynamically stripping sentences and facts from all content sources, and recombining them. the computer writes a new story for every user'
5'55": 'in 2011 the slumbering fourth estate awakes to make its first and final stand. the new york times company sues googlezon, claiming that the company's fact-stripping robots are a violation of copyright law. the case goes all the way to the supreme court'
they didn't get the details exactly right, but overall the accuracy is astounding
however, that may be a hyperstition artifact in this timeline
https://en.wikipedia.org/wiki/EPIC_2014 (i thought epic 2014 might be the only flash video to hae a wikipedia article about it, but then i looked and found five others)
However, the suit goes far beyond claiming that such copying violates their copyright: "Unauthorized copying of Times Works without payment to train LLMs is a substitutive use that is not justified by any transformative purpose."
This is a strong claim that just downloading articles into training data is what violates the copyright. That GTP outputs verbatim copies is a red herring. Hopefully the judge(s) will notice and direct focus on the interesting, high-stakes, and murky legal issues raised when we ask: What about a model can (or can't) be "transformative"?
I can't take NY Times articles, translate them into Spanish, and then sell the translations under fair use, even though clearly I've transformed the original article content.
Suppose I’m selling subscriptions to the New Jersey Times, a site which simply downloads New York Times articles and passes them through an autoencoder with some random noise. It serves the exact same purpose as the New York Times website, except I make the money. Is that fair use?
They transformed the weights.
Just like reading the article transforms yours.
As for verbatim reproduction, I'm pretty sure brains are capable of reproducing song lyrics, musical melodies, common symbols ("cool S"), and lots of other things verbatim too.
Those quotes from Dr. King's speech that you remember are copyrighted, you know?
They should be.
Times change. We're industrializing information creation and consumption (the latter is mostly here already), and we can't be stuck in the old copyright regime. It'll be useless in very short order.
All this road bump will do will give the giant megacorps time to ink deals, solidify their lead, and trounce open source. Twenty years on, the pace of content creation will be as rapid as thought itself and we'll kick ourselves for cementing their lead.
This is a transitional period between two wildly different worlds.
But nobody would do that, because ChatGPT is a really shitty way to read NYT articles (it's stale, it can't reliably reproduce them, etc.). All that is valuable about it is the way that it transforms and operates on that data in conjunction with all the other data that it has.
The real world use of ChatGPT is very transformative, even if you can trick it into behaving in ways that are not. If the courts act intelligently they should at least weigh that as part of their decision.
Suppose I start a service called “EastlawAI” by downloading the Westlaw database and hiring a team of comedians to write very funny lawyer jokes.
I take Westlaw cases and lawyer jokes and feed them to my autoencoder. I also learn a mapping from user queries to decoder inputs.
I sell an API and advertise it to startups as capable of answering any legal question in a funny way. Another company comes along with an API to make the output less funny.
Have I created a competitor to Westlaw by copying Westlaw’s works for their original expressive purpose and exposing it as an intermediary? Or have I simply trained the world’s most informative lawyer joke generator that some of my customers happen to use for legal analysis by layering other tools atop my output?
Did I need to download Westlaw cases to make my lawyer joke generator? Are the jokes a fair-use smokescreen for repackaging commercially valuable copyrighted data? Does my joke generator impact Westlaw in the market? Depends, right?
To be clear, whether the use of the original work is transformative is one key consideration within one of the four prongs of fair use. The prong "purpose and character of the use" can be fulfilled by other conditions [1]. For example, using the original work within a classroom for education purposes is not transformative, but can fulfill the same "purpose and character of the use" prong. Whether the use is for profit and to which extent are other considerations within that prong. A profit purpose doesn't automatically fail the purpose prong, and a non-profit purpose doesn't automatically pass the purpose prong.
[1] https://en.wikipedia.org/wiki/Fair_use#1._Purpose_and_charac...
This is not a RLHF problem. What I was expecting them to do is to keep a bloom filter of ngrams for known copyrighted content, such as enumerating all sets of n=7 consecutive words in an article, and validate against it. The model would only output at maximum n-1 words that look verbatim from the source.
But this will blow up in their face. Let's see:
- AI companies will start investing much more in content attribution
- The new content attribution tools will be applied on all human written articles as well, because anyone could be using GPT in secret
- Then people will start seeing a chilling effect on creativity
- We must also check NYT against all the other sources, not everything the write is original
- Paraphrasing n=7 words (and quite a few more) within a sentence can easily be fair use.
- As n gets big, the bloom filter has to also.
If/when attribution is solved for LLMs (and not fake attribution like from Bing or Perplexity) then creators can be compensated when their works are used in AI outputs. If compensation is high enough this can greatly incentivize creativity, perhaps to the point of realizing "free culture" visions from the late 90s.
I tested this 6-gram "it won't find anything matching exactly", no match. Almost anything we write has never been said exactly like that before.
This approach is probably inadequate. In my line of (NLP) research I find many things have been said exactly many, many times over.
You can try this out yourself by grouping and counting strings using the many publically available Bigquery corpora for various substring lengths and offsets, e.g. [0-16]; [0-32]; [0-64] substring lengths at different offsets.
Who pays the compensation? If it's the user, why wouldn't they just buy the authors work directly? Why go through the LLM middleman?
If it's the user, why wouldn't they just buy the DVDs directly? Why go through the Netflix middleman?
A retort to this would be that both NYT and ChatGPT are on the internet, so it's no added fuss of hopping in my car, driving to Walmart, and picking up a DVD case. My response to it would be that both the LLM and Netflix are content aggregators to the user. I can read the NYT, or I can read the NYT summary on ChatGPT and ask it for life advice with my pet hamster, or ask it how to reverse a linked list in bash.
Then there's the issue that however you credit attribution, it creates a game of enshittified content creation with the aim of being attributed as often as possible, regardless of whether the content really offered anything that wasn't out there already.
Specifically, the NYT examples all seem to be cases where they asked the AI to repeat their articles verbatim? So they ask it to violate copyright and because it's a helpful bot with a good memory, it does so.
Solution: teach the model to refuse requests to repeat articles verbatim. It's easily capable of recognizing when it's being asked to do that. And that's exactly what OpenAI have now done.
So the direct problem the NYT is complaining about - a paywall bypass - is already rectified. Now it would seem to me like the case is quite weak. They could demand OpenAI pay them damages for the time ChatGPT wasn't refusing, but wouldn't they have to prove damages actually happened? It seems unlikely many people used ChatGPT as a paywall bypass for the NYT specifically in the past year. It only knows old articles. OpenAI could be ordered to search their logs for cases where this happened, for example, and then the NYT could be ordered to show their working for the value of displaying a single old article to a non-subscriber, and from that damages could be computed. But it wouldn't be a lot.
That's presumably why the case goes further and argues that OpenAI is in violation even when it isn't repeating text verbatim. That's the only way the NYT can get any significant money out of this situation.
But this case seems much weaker to me. Beyond all the obvious human analogies, there is precedent in the case of search engines where they crawl - and the NYT let them crawl - specifically to enable the creation of a derived data structure. Search engine indexes are understood to be fair use, and they actually do repeat parts of the page verbatim in their snippets. Google once even showed cached versions of whole pages. And browser makers all allow extensions in their stores that strip ads and bypass paywalls, and the NYT hasn't sued them over that either.
This demonstrates that no, the NN actually does contain the full articles, copied into the NN. Do you think any normal person would get away with copying MS windows by e.g. zipping it together with some other OS on the same medium. Why should we let OpenAI get away with this?
> Why should we let OpenAI get away with this?
IP rights, like other private property rights, are a compromise between creators and consumers. What "should" be the case is essentially an argument about what balance creates the best overall outcomes. LLMs, for now, require large amounts of text to train, so the question is one of whether we want LLMs to exist or not. That's really a question for Congress and not the courts, but it'll be decided in the courts first.
The entity which owns ChatGPT is apparently maintaining a copy of the entirety of the New York Times archive within the ChatGPT knowledge base. That they extract some fair use snippets (they would claim) from it would still be fruit of a poisoned tree, no?
(disclaimer: I'm pro AI, anti copyright, especially anti elitist NY Times; but pro rule of law)
Your creative work does deserve at least some period of exclusive rights for you. Definitely not so much that your grandchildren get to quibble about it well into retirement. But also whatever the number 3 or 4 most valuable company in the world doesn’t get to scrape your content daily to repackage and sell as intelligent systems.
Here's a thing though: for 99%+ of that content, being turned into feedstock for ML model training is about the only valuable thing that came of its existence.
If it were not for world-ending danger of too smart an AI being developed too quickly, I'd vote for exempting ML training from copyright altogether, today - it's hard to overstate just how much more useful any copyrighted content is for society as LLM training data, than as whatever it was created for originally.
Note that I don't see any major problem if only articles that were, say, more than 5 or 10 years old were being used. I don't think the current length of copyright makes any sense. But there is a big difference from last year's archive vs today's news.
This can be further coupled with search - use GPT to look at multiple sources at once, and report. It's what humans do as well, we read the same news in different sources to get a more balanced take. Maybe they have contradictions, maybe they have inaccuracies, biases. We could keep that analysis for training models. This would also improve the training set.
LLMs are arguably compressed data archives with weird algorithms. The fact that they will regularly regurgitate verbatim quotes of training data is evidence of this, as are the guardrails that try to prevent this.
The second piece of evidence is this paper explained here https://www.hendrik-erz.de/post/why-gzip-just-beat-a-large-l... where instead of an LLM researchers used gzip compressed data as a model and it even beat trained LLMs.
AI is a bit of a black box, but that doesn’t protect the operators of black boxes from rights violation suits. You can’t make a database of scraped copyrighted data and patented that querying that data is fair use.
There needs to be law made here and the law just isn’t going to be “everybody can copy everything for free as long as it’s for model training”.
Licensing will have to be worked out, actual laws and not just case law needs to be written. I have a lot of sympathy for lots of leeway for the open source researchers and hackers doing things… but not so much for Microsoft and Microsoft sponsored openai.
Further, the evidence presented by NYT in the lawsuit could be hard to reproduce. I tried multiple prompts on multiple versions of GPT-4 APIs but still could not get GPT-4 to reproduce NYT articles exactly. NYT might as well tried to let GPT-4 reproduce 100,000 articles and only found a few cases where GPT-4 actually recited the whole article. In that case OpenAI might as well be arguing that this is only a rare bug and avoid losing the lawsuit in a massive way.
OpenAI has created a $100bn company on this transfer. The Times may have an interest in a material fraction of that wealth.
The Times almost certainly wants its own LLM. I could see them striking a consortium agreement with other newspapers more easily than OpenAI.
OpenAI alone has a market cap that'd allow it to buy about as large a proportion of publishers of newspapers and books as they'd be allowed before competition watchdogs will start refusing consent.
Put another way:
If I was a VC with deep pockets investing in AI at this point, I'd hedge by starting to buy strategic stakes in media companies.
I'm not sure how your proposal would actually work. To recognize plagiarism during inference it needs to memorize harder.
Kinda funny if it works though. We'd first train them to copy their training data verbatim, then train them not to.
That is how it works, right? They're trained to copy their training data verbatim because that's the loss function. It's just that they're given so much data that we don't expect this to be possible for most of the training data given the parameter count.
[1] https://bair.berkeley.edu/blog/2019/08/13/memorization/
One thing you might do is use a full-text search database of the entire training data. If part of ChatGPT response is directly copied, give it the assignment of "please paraphrase this" and substitute the paraphrase into the response. This might slow ChatGPT down a lot - but it might not, I think an LLM is actually more computationally expensive than a full-text search by a lot.
Production open access LLMs do probably need a front-end filter with a fine tuned RAG model that identifies and prevents spitting out copyrighted material. I fully support this.
But we shouldn't be preventing the development of a technology that in 99.99% of usecases isn't doing that and can used for everything from diagnosing medical issues to letting coma patients communicate with an EEG to improving self-driving car algorithms because some random content producer's works were a drop in the ocean of content used to learn relationships between words and concepts.
The edge cases where a model is rarely capable of reproducing training data don't reflect infringement of training but of use. If a writer learns to write well from a source is that infringement? Or is it when they then write exactly what was in the source that it becomes infringement?
Additionally, now that we can use LLMs to read brain scans and have been moving towards biological computing, should we start to consider copying of material to the hippocampus a violation of the DMCA?
Isn't that in tension with the basic idea of an LLM of predicting the next token? How do you achieve that while never getting close enough to plagiarism?
I feel like the NYTimes is asking for deletion as a negotiation tactic to force OpenAI to give them enough money to pay for their journalism (I am not sure who would subscribe to NYTimes if you can get as much through OpenAI, but I am open to registering extra to pay for their work).
It's the other way around. There is no infringement if the model output is not substantially similar to a work in the training set [1]:
> To win a claim of copyright infringement in civil or criminal court, a plaintiff must show he or she owns a valid copyright, the defendant actually copied the work, and the level of copying amounts to misappropriation.
The questions are, which parties should bear liability when the model creates infringing outputs, and how should that liability be split among the parties? Given that getting an infringing output likely requires the prompt to reference an existing work (which is what's happening in the article), an author of a work, an element in an existing work, or a characteristic/style strongly associated with certain works/authors, I believe that the user who makes the prompt should bear most of the liability should the user choose to publish an infringing output in a way that doesn't fall under fair use. (AI companies should not be publishing model outputs by default.)
[1] https://en.wikipedia.org/wiki/Substantial_similarity#Substan...
Its true that OpenAI will defend the wholesale copying into the training set by arguing that the transformative purpose of the next use reaches back and renders that copying fair use, but while that's clearly the dominant position of the AI industry, and it definitely seems compatible with the Cobstitutional purpose of Fair Use (while currently statutory, the statutory provision is codification of Constitutional case law), it is a novel fair use argument.
NY Times is suing because of both the model outputs and the existence of the training set. But infringement in the training set doesn't necessarily mean that the model infringes. Why? Because of the substantial similarity requirement. But first, I'll address the training set.
For articles that a person obtains through legal methods (like buying subscriptions) but doesn't then republish, storing copies of those articles is analogous to recording a legally accessed television show (time-shifting), which generally is fair use. Currently, no court has ruled that "analogous to time-shifting" is good enough for the time-shifting precedent to apply, but I think the difference is not significant. The same applies to companies. Companies are not literally people, but there isn't a reason for the time-shifting precedent to not apply to companies.
What about the articles that OpenAI obtained through illegal methods? Then the very act of obtaining those articles would be illegal. The training set contains those copies, so NY Times can sue to make OpenAI delete those copies and pay damages. But it's not trivially obvious that a GPT model is a copy of any works or contains copied expression of the any works in the training set; the weights that make up the model represent millions of works, it's not trivially obvious that the model contains something substantially similar to the expression in a work in the training set. Therefore, it's not trivially obvious that infringement with respect to the training set amounts to infringement with respect to the model made from the training set. If OpenAI obtained NY Times articles through illegal means, then making OpenAI delete the training set would be reasonable, but the model is a separate matter.
As long as the model doesn't contain copied expression and the weights can't be reversed into something substantially similar to expression in the existing works, then what matters is the output of the model.
If a user gives a prompt which contains no reference to an existing NY Times author, work, or a strongly associated characteristic/style, then do OpenAI's models produce outputs substantially similar to expression in the existing works? If not, then OpenAI shouldn't be liable for infringing works, because the infringing works result from the user's prompts. If my premise is false, then my conclusion falls apart. But if my premise is true, then at most I would admit that OpenAI has a limited burden to prevent users from giving those prompts.
The thing about you claim, "Just learn to recognize and punish plagiarism via RLHF" is that we've had an endless series of prompt exploits as well as unprompted leakage and these demonstrate that an LLM just doesn't have fixed border between its training data and its output. This will it basically impossible for OpenAI to say "we can logically guarantee ChatGPT won't serve your data freely to anyone".
Also lying on source materials (e.g. telling students that some respected historian denies the Holocaust happened, when it's obviously not the case) is not "teaching" - it's defamation, and the NYT is absolutely right to pursue that angle too.
Using LLMs as general-purpose search engines is a minefield, I would not be surprised if the practice disappeared in the next 20 years. Obviously the tech is here to stay, there is no problem when it's applied to augmenting niche work; but as a Google replacement, it has so many issues
Incorrect. Educational use helps satisfy one of tests for fair use. Teachers can, in many cases, photocopy copyrighted work without infringing on that copyright.
https://youtu.be/j_UoACEUZqA
If I go fishing, the regulations I have to comply with are very light because the effect I have on the environment is minimal. The regulations for an industrial fishing barge are rightfully very different, even if the end result is the same fish on your plate.
https://www.goodreads.com/quotes/21810-it-is-difficult-to-ge...
You have no idea what you’re talking about huh?
In fact all the demonstrations in the lawsuit PDF were intentionally angling for reproducing copyrighted content. They had to push the model to do it. That won't happen unless users deliberately ask for it. It won't happen en-masse.
Boo hoo they had to push it. That was never the problem with these bullshit nozzles. The issue is they put that stuff in the training set in the first place. If you can't be honest about that then I have no interest in debating this with you.
Discussion here: https://news.ycombinator.com/item?id=38781941
Then LLMs would be distributed only via torrents, like most copyright infringing media.
Why?
The LLM genie is out of the bottle: an unfavorable court ruling in a single country isn't going to stuff it back in.
On the other hand, if LLM are used to "launder" copyright content and, accepting the premises of copyright law, this has the effect of reducing incentives to do creative work, that has obvious negative implications for economic productivity.
Assuming this is in good faith: the ability to write code, documentation, and tests is absolutely a productivity enhancer to an existing programmer. The code snippets from a dedicated tool like copilot are of very usable quality if you're using a popular language like Python or JS.
Loading data to which you have no rights over into your software is legally perilous, yes.
It's as easy as simply asking for and receiving permission from the data's rightsholders (which might require exchange of coin) to make it not legally perilous.
I suspect it wouldn't be too hard to convince the EU though, the EU has an history of giving up rights and markets to big copyright holders even if that hurts the local companies.
LLM's will become more expensive and less attractive as money printers, this will screw with the business models of the direct provision folks like OpenAI, MS and Google, MS and Google will only shed tears for money spent while OpenAI will just not have as good an income stream until they think of something new.
I'm sure that's what they want, but I'm not sure that's what the outcome will be. What if they want to charge a prohibitive amount of money for their content?
I think Spotify vs Napster is a good example, content creators in news (the Journalists) are already in a hard place (vs. successful rock stars preinternet) I think that the news providers are rather like the music lables.
So you will use a Chinese AI that spies on you, or you will use some shady service from a shady country (that will play cat and mouse like torrent sites).. or most likely you will run your own model when you are computer literate and no model if you are not.
Actually most models are so lobotomized allready that probably better to run your own, as long as you have a good enough computer.
No need to emigrate!
https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec20...
From page 30 and onwards has some fairly clear examples on how ChatGPT has an (internal) copy of copyrighted material which it will recite verbatim.
Essentially if you copy a lot of copyrighted material into a blob and then apply some sort of destructive compression to it. How destructive would that compression have to be for the copyright no longer to hold? My guess it would have to be a lot.
As I see it the closeness of OpenAI may be what saves it. OpenAI could filter and block copyrighted material from the LLM from leaving the web interface using some straight forward matching mechanism against the copyrighted part of the data set ChatGPT has been trained on. Whereas open source projects trained on the same data set would be left with the much harder task of removing the copyrighted material from the LLM itself.
I imagine the goal is closer to "enough that no one notices we stole it", either in a way that it's not easily discoverable or even when directly analyzed there's enough plausible deniability to scrape by.
It makes it difficult for me to ascertain whether it is repeating from it's training data, or they committed the same mistake as the OP article of using Copilot, which ends up googling(binging?) the article first, before replying.
A web-crawled LLM that lived within the same constraints would be a search engine under another name, with a slightly different presentation style. If it starts spitting out entire articles without citation, that's not acceptable.
If I use ChatGPT as a research tool, as long as it lives within the same parameters that I have to live within, I don't see a problem with its education/learning.
I understand that the NYTimes would like a slice of anything that comes out of the GPT but I'm talking about what seems reasonable. People who share their copyrighted material do not own all of the thinking that comes out of it; they own that expression of it, that is all.
Will AI destroy the economics of "writing" the way the web has killed newspapers? perhaps, perhaps we'll all benefit from and need a new model, but killing the new to keep the old on life support is not the way.
I'm not saying LLMs are by default, illegal. All I'm saying is that there is some merit to why NYT and content companies want a piece of the pie and think they deserve it.
Let's say I host these, offer some indexing on it, and rewrite articles. Something like, summarise all articles on US-UK relationships over past 5 years. I charge money for it, and all I pay NYT is a monthly subscription fee. To keep things simple, let's say I never regurgitate chunks of verbatim NYT articles, maybe quite short snippets.
Is that fair use? IANAL, but doesn't sound like it. Typically I can't take a personal "tier" of a product and charge 3rd parties for derivatives of it. Say like VS Code.
A sibling comment mentions search engines. I think there's a big difference. A search engine doesn't replace the source, not at all. Rather it points me at it, and offers me the opportunity to pay for the article. Whereas either this or an LLM uses NYT content as an alternative to actually paying for an NYT subscription.
But then what do I know...
>Is that fair use? IANAL, but doesn't sound like it.
If you pay someone to do the summarisation for you, then you publish the content and charge a fee for it, you're the one liable, not the person you paid to summarise it for you. Similarly if you ask GPT to do it for you, then publish it, you're liable for what you publish; GPT is just a summarisation tool.
At some level it becomes a subversion of NYTs fees. First, say I subscribe and simply host the articles verbatim, for a fee. Clearly, that's not right.
Suppose I change some spelling or word order, or use a synonym or two. That's still not ok.
And if I substantially paraphrase the articles? I guess this is the relevant case. This is kind of what LLMs do. And also feels like not fair use.
That's not what OpenAI is doing; it's not selling summarised articles as a service. Your example is a false equivalence.
>This is kind of what LLMs do. And also feels like not fair use
An LLM doesn't do this unless you ask it to. And if you then take that output and publish it as your own, you're breaching the copyright, not OpenAI.
In this case, OpenAI is violating copyright by modifying, reproducing and distributing copyrighted content to its customer.
I read a NYT article, then summarize it into a link title for reddit. Reddit then republishes the summary to all of its users.
So, if the summaries are derived works and not covered by fair use, then both you and the summarizee are separately breaking the NYT's copyrights. Otherwise, if this is covered by fair use, then you are both in the clear.
Finally, GPT is not "a summarization tool" in this case. If you provide a copy of a NYT article as a prompt and then ask for summarization, then yes, it is clear that GPT is not doing anything wrong, even if it spits out the exact same text. But if you simply ask for a summary of a specific article by, say, just name and date, and you get a copy of it, it's clear that GPT is storing the original data in some way, and thus it has copied the NYT's protected works without permission.
In this particular case they were using it via Bing, which actively did a HTTP request to the particular article to extract the content. So GPT hadn't memorised it verbatim, instead it fetched it, much like a human using a search engine would.
Additionally, even the use through Copilot is very debatable. They are not returning the NYT link, which requires a subscription, they are returning the contents of it even to non-subscribers. And they are doing this in a commercial product, not a non profit like the Internet Archive, which has some arguments for fair use.
Sometimes they're so overfit that the compression isn't even lossy, and the data is encoded verbatim in the NN.
But I disagree with the underlying assumption that you can anthropomorphize LLMs. Gradient descent and backpropagation don't take place in the brain. LLMs "learn" in the same way that Excel sheets "learn".
Humans are living beings with needs and rights. A person being able to legally squat in a home doesn't mean that a drone occupying property for some amount of time also has squatter's rights, even though you could easily and affordably automate and scale the deployment of drones to live and hide away on properties long enough to attain rights regarding properties all over the country.
I await the HN ban with fear..
[1] I'm not even doing referencing - so I am surely an LLM.
but, more importantly, OpenAI can also be sued for tortious interference? (basically the civil equivalent of accessory)
That's function of the legal system, not of the technology. If tomorrow someone made a perfect dolphin-Esperanto translator and proved Dolphins were as smart as humans, you still can't sue a dolphin until the legal system says so.
Not exactly, no, but the 'neurons that fire together wire together' way of learning has a pretty similar effect.
> LLMs "learn" in the same way that Excel sheets "learn".
I've never seen an excel sheet do anything like backpropagation.
Not strictly in the sense you mentioned (assuming that you mean "by themselves") but people may find [1] and [2] interesting.
[1] https://pub.towardsai.net/building-a-neural-network-with-bac...
[2] https://towardsdatascience.com/demystifying-feed-forward-and...
Backprop doesn't happen in us, but I think our neurones still do gradient descent – synapses that fire together, wire together.
And ultimately, at the deepest level we can analyse, our brains' atoms are doing quantum field diffusion equations, which you can also do in an Excel spreadsheet, so that kind of reductionism doesn't help either.
> Humans are living beings with needs and rights. A person being able to legally squat in a home doesn't mean that a drone occupying property for some amount of time also has squatter's rights, even though you could easily and affordably automate and scale the deployment of drones to live and hide away on properties long enough to attain rights regarding properties all over the country.
Yes, but we can also do tissue cultures and crude bioprinting, so it's a very foreseeable future where exactly the same argument will also be true for living organisms rather than digital minds.
We need to figure out what the deeper rules are that lead to the status quo, not merely mimic the superficial result. The latter is how cargo cults function.
Sure, that's an interesting path of inquiry, and one should be free to understand themselves as being no different than a machine if they desire.
But the objective of laws is the benefit of (at least some) humans, not machines covered in lab grown tissue. The process of being human is a big part of what makes us human.
I think you're misapprehending — I mean an entity fully 3D printed out of tissue, no machinery (unless you're counting all biology as machinery, but I think you're not doing that).
I recon bio-printing is now where home computing was in the Apple 1 era, so this is a way off, but it's foreseeable.
> The process of being human is a big part of what makes us human.
Mmm. How much has that process that changed since the ancient world?
How do you recon that, Apple 1 was Turing complete. We haven't printed life, that would be a tremendous accomplishment.
I think we're closer to Edison inventing a lightbulb as a step to computers being possible. Printing a conscious thing, at all, would be like the transistor. An Apple 1 analogue wouldn't be likely because of the terrible ethics of a "shitty" printed human.
Sure we have, and in multiple different senses.
The ones which matters here are cell culture, which is nowhere near the fanciest bar that's been surpassed in this field, and tissue culture which is somewhat harder but the reason why I recon it's at the Apple 1 level is that a small number of experimentalists are messing around with it using expensive equipment that you can technically buy at home but you need to be well trained to actually use, for example:
https://youtu.be/Z_ZGq8Tah0k?si=u6bBatjuSWcyNYJ3
And, more broadly, there's bioprinting as a research etc. field:
https://en.wikipedia.org/wiki/3D_bioprinting
And here's a TED talk from ten years ago where they demoed an early research 3D printed kidney on stage:
https://youtu.be/bX3C201O4MA?t=11m46s
No. That isn't printing life, that is taking already living cells, priming and transforming them into something useful. Regardless, I'd count it if we could make an entire living organism this way, but we cant. Creating a working organ is no doubt amazing, and proof that this technology is worth pursuing, but it isn't "printing life" any more than producing life saving drugs is.
In your example you are talking about being able to bioprint a person(they have to be a person to have that right) to squat a property. Bio printing an organ isn't an example of that, it's not even close. Saying that we are anywhere near being able to print a human to squat a property is pretty ridiculous.
Which is absolutely sufficient for the usage I described upthread. In fact, I'd go so far as to say it's mandatory for the point I was making, as — fun though bio-printed werewolves, dragons, and fae would be — my point only works if you get humans out of the process rather than some other species. A bioprinted horse is probably slightly harder than a bioprinted human, but the latter isn't getting any squatting rights.
I could've linked to work on synthetic genomes and nucleotides to give evidence for lower-level creation of live, but they don't matter for the same reason:
My point is that there's a pathway heading off into the distance, and somewhere in the distance but before the horizon can be found bio-printed humans with all the same moral issues we're now just beginning to taking seriously thanks to AI being conversational, and if we had something completely customised, that's cool and all, but it doesn't make anyone go "oh, they're people" the way a humanoid body with human DNA getting off a table saying "hello, nice to meet you" does.
> In your example you are talking about being able to bioprint a person(they have to be a person to have that right) to squat a property. Bio printing an organ isn't an example of that, it's not even close. Saying that we are anywhere near being able to print a human to squat a property is pretty ridiculous.
I wrote "an entity fully 3D printed out of tissue […] is a way off, but it's foreseeable" and compared bio-printing today to a nearly 50 year old computer, and one of my references was a link to a youtube channel where someone is attempting to do a small-scale prototype thing along these lines with a handful of organs made from mouse cells grown in his own lab (and mouse cells rather than human because of the disease risk not because something magic happens with human cells). You're mixing up what I think is foreseeable with what I say already exists, and using the nonexistence of what I think can be foreseen to argue against what does exist.
No! Hebbian learning is categorically NOT gradient based learning. Hebbian update rules are local and not the gradient of any function.
Cortical learning is so vastly different from how artificial neural networks “learn” they cannot even begin to be meaningfully compared mathematically. Hebbian learning is not optimization and backprop is not local learning.
Part of the problem of these discussions is a bunch of clueless people talking with authority.
I have to keep reminding myself that outside of my own speciality, ChatGPT knows more than me despite its weaknesses, so I bet ChatGPT knows more about Hebbian learning than I do.
I'll look into that more.
most of the world disagrees with this view, and that means they will create the AI that wins.
You misunderstood me. I was talking about something more fundamental.
Understanding is data compression. They are the same thing. Learning patterns, building mental models, creating abstractions, generalizing, gaining intuition/a feel for something - all the things humans engage in as part of learning and understanding the world - are all acts of lossy data compression.
Anyone got more details on this?
Superficially it sounds like total BS; a highly compressed zip file does not exhibit any characteristics of learning.
Algorithmically derived highly compressed video streams do not exhibit characteristics of learning.
?
I’ve vaguely heard the learning can be considered to exhibit the characteristics of compression in that understanding of content (eg. segmentation of video content resulting in more highly compressed videos) can lead to better compression schemes.
…but saying you can “do a with b” and “a and b are fundamentally the same thing” seems like a leap…?
It seems self evident you can have compression without comprehension.
An LLM has limited parameters. If an LLM had infinite parameters it could just memorize the results of every single addition question in existence and could not claim to have understood anything. Because it has finite parameters, if an LLM wants to get a lower loss on all addition questions, it needs to come up with a general algorithm to perform addition. Indeed, Neel Nanda trained a transformer to do addition mod 113 on relatively few examples, and it eventually learned some cursed Fourier transform mumbo jumbo to get 0 loss https://twitter.com/robertskmiles/status/1663534255249453056.
And the fact it has developed this "understanding" as an ability to learn a general pattern in the training data enables it to compress. I claim that the number of bits required to encode the general algorithm is fewer than the number of bits required to memorize every single example. If it weren't then the transformer would simply memorize every single example. But if it doesn't have space then it is forced to try to compress by developing a general model.
And the ability to compress enables you to construct a language model. Essentially, the more things compress, the higher the likelihood you assign them. Given a sequence of tokens say "the cat sat on the", we should expect "the cat sat on the mat" to compress into fewer bits than "the cat sat on the door". This is because the latter is far more common and intuitively more common sequences should compress more. You can then look at the number of bits used for every single choice of token following "the cat sat on the" and thus develop a probability distribution for the next token. The exact details of this I'm unclear on. https://www.hendrik-erz.de/post/why-gzip-just-beat-a-large-l... this gives a good summary.
I fundamentally disagree. That's not some established fact, just a narrative used by those who wish to plagiarize using "AI".
Our collective human limitations(physical, mental and temporal) are sort of invisible implicit rules that we all follow in one way or the other. If an entity is not bound by those rules then I don't see why that entity should be treated the same as a human.
Companies already make this differentiation.
For example take captcha and bot detection. Some of the heuristics are based on inherent human limitations like response time, click time, mouse acceleration etc.
I doubt youtube or any other streaming service will be happy if you want to stream all their videos to train a hypothetical human like AI(which views and prepares notes like a human) at a hugely accelerated speed compared to a regular human. You can guess how quickly they will cite fair usage policies.
What I want to say is there are fundamental differences between a human and an AI. So, we should not be quick to dismiss any concerns just because AI can "mimic" humans in certain areas.
Humans have rights, software tools don’t.
If you grant an LLM the full set of human rights, then it can consume information, regurgitate copyrighted works, and use it to generate money for itself. However, considering blatantly obvious theft as “homage” goes hand in hand with free will, agency, being in control of yourself, not being enslaved and abused, etc. Pondering various scenarios along those lines really gets to the heart of why an LLM is so very much not a human, and how subjecting it to the same treatment as humans is a ridiculous notion.
If you don’t grant LLM human rights, then ClosedAI’s stance is basically that pirating works is OK because they pass them through a black box of if conditions and it leads to results that they can monetize. That’s such a solid argument, it’ll surely play well in the court of law.
Training data is not an “LLM does it”; first because “it” here is not “learning” or understanding in human sense (otherwise you would have to presume that an LLM is a human), and second because a software tool doesn’t have agency and it’s really just Microsoft using a tool based on copyrighted works to generate profit.
What I expect to happen is whoever has the most influence and power will get what they want and we'll end up raising a generation with the implicit understanding of "that's just how things are," natural order, truth, reality, and all that jazz.
The only thing that ever changes outcomes is if the contradiction status quo is incapable of being managed.
Here's an article from November 2023 that discusses this:
https://not-just-memorization.github.io/extracting-training-...
Can't you, though? I'd thought in general, it's a very important for the market to be able to do just that, otherwise everything gets gummed up in webs of exclusive contractual dependencies between established companies.
Typically providers of online databases go to some effort to stop people from sharing logins. Even from that point or view, I can imagine scraping articles and providing paraphrases of it for a fee is fishy.
All I'm saying, to some people it's obvious that the whole LLM on scraped Internet is fair use, to me it is not obvious.
Seems like the "problem" is that NYT etc gives privileged access to search engines for indexing their content, but then get upset when snippets of the indexed content is being shown to users without the users having to fight the paywall or whatever.
This article also claims that the screenshot is coming from ChatGPT when it clearly is not.
I'm not sure the problem goes away simply if the LLM in question (or any other one) gets some "no verbose regurgitation" filter.
Edit: it looks like you've unfortunately been breaking the site guidelines quite a bit lately. Can you please review them and stick to the intended use of the site? We'd appreciate it.
https://news.ycombinator.com/newsguidelines.html
It’s not clear to me where the line is.
If the verbatim examples that have been going around are true, that’s bad. I’d love to know more details around it — prompts used, whether that’s an old model, etc. This seems like plagiarism more than anything.
Yielding verbatim snippets of copyrighted content is a problem for OpenAI though.
So the demand to destroy those databases seems very dubious to me.
Of course later violating fair use is another issue.
As always, the answer is.. "it depends". I guess it depends mostly on the jurisdiction that applies to you. "Fair use" can have rather different legal meaning (or not exist at all) in different countries.
This is demonstrably wrong. Many countries have both freedoms, albeit some have less strong protection than others.
They said "Most other countries" and you replied with "Many countries"
"Many" does not necessary include "most" but "most" does include "many".
No, it doesn't. If a set is of sufficiently low cardinality, “most” (in extreme cases, even “all”) of the set may not be “many”.
Most-all, in fact—Catholic Presidents of the United States have been Democrats. But it is not the case that many Catholic Presidents have been Democrats.
Most women to have served on the US Supreme Court did so only after its first 200 years. But, again, there were not many women who served on the Supreme Court only after its first 200 years.
I replace the following sentence from my previous comment:
> Most other countries don't have freedom of expression and freedom of the press, so copyright law in a different country usually lacks a unifying exception test like fair use to supplement the specific enumerated exceptions.
with the following:
Copyright law in most countries usually lacks a unifying exception test like fair use to supplement the specific enumerated exceptions in each respective country.
The rest of my previous comment remains the same.
I think you’re confusing terms of service and copyright. IANAL but what you describe sounds exactly like fair use to me, irrespective of how much you are paying NYT.
I think people severely underestimate how much they've grown accustomed to this information being freely available. It's easy to say "Well it shouldn't be available with ChatGPT," but if we actually put everything back behind a paywall and stopped people from doing things like writing blogs or newsletters that summarize the news, people here would get angry very fast.
Google has been accused for years of replacing sources with their "One Box"--the big answers at the top of the page, which are usually pulled from or corroborated by search results. They don't want you to leave the search results page (where the ads are).
Instead, I think they're paying for this:
https://enterprise.wikimedia.com/
Not only that, look at a few news articles from Tier 2 and down publications, and you'll realize that almost all of them are directly sourced from NYT and others. They'll say "so and so happened, according to The Times" (and usually link the article there)
Just like during the pandemic how everyone became an epidemiologist, suddenly everyone's a copyright lawyer. I'll just dispute your assertion by saying:
1. Questions of fair use are famously gray, and anyone who declares something as "entirely fair use", with no caveats, is nearly always wrong except for the must obvious cases, which the given example is most definitely not. A judge has wide latitude in determining fair use.
2. People should familiarize themselves with the four factors of fair use determination. In particular, if a work is purely derivative of a source work and substantially negatively impacts the market for the original work, it's very likely to not be considered fair use.
A great overview is https://fairuse.stanford.edu/overview/fair-use/four-factors/
Roll back 20+ years ago on Slashdot and you'll see the exact same thing.
Copyright has been a hot button issue on the internet for decades. People end up thinking (rightly or wrongly) that they understand it without being a lawyer.
It's very possible that the example provided above is an example of fair use in some country, and that the website offering that service could be hosted there.
Quite literally, not even the lawyers or courts understand it. This is very much a "learn as you go" exercise for humanity in general at this point in time.
I only see this phenomenon speeding up. Strange times.
This is just a felony contempt of business model issue. Computers invalidated their business models and they're doing everything they possibly can to hang on for dear life. Society needs to move on already.
Books are just strings of letters, yet copyright has still been useful to increase the volume and utility of books.
All that said, I do find the life+70y an absurdly long time.
What do you propose is the business model for artists in the absence of copyright?
We must strengthen these business models that don't depend on artificial scarcity because this number selling nonsense was over the second computers were invented. It's as dumb as asserting that you need permission to use memcpy or the mov CPU instruction.
In the US, the original copyright length was 14 years, and then 28, and eventually the lifetime of the author plus 70 years. I think the intent of the law is economically justified, but the current length is outrageous.
This is how art worked for millenia; someone commissions a chapel roof painting, someone commissions a concerto, someone commissions a statue, someone buys a chair, etc.
Artists still do this today, and there is no issue determining value beforehand. Artists list their commission prices, or their hourly costs, etc. This is a perfectly normal thing that happens everyday.
It seems to have been invented by laywers, for lawyers. Nobody else really benefits as much as they do. The whole entirety of society vs. a single profession of dubious morality.
meanwhile, most tech is moving towards subscriptions?
Art is getting paid "non-greedily". People buy a song or art piece, and then people 10 years later buy a song or art piece. That's not one person paying twice for the same song, it's two people buying the same thing.
If people still value that art for that price later, I don't see how this is a "greedy" thing. is art magically supposed to turn open source CC0 after 5 years? Tech sure doesnt work like that.
But ok so I'm a young musician. Nobody's heard of me and nobody wants to commission a concert or album. What do I do? Quit?
@Vicinity9635 It's not just lawyer greed, it creates economic fairness by preventing others from profiting from your creative work.
In the current system, artists might work for many years on a single work, or work many years perfecting their craft before anyone wants to pay for their work. Copyright gives them a way to earn money in the future that compensates them for the work they did in the past. It incentivizes creativity. Don't get me wrong, I don't think copyright is perfect, but you really ought to think more about the system you're proposing, because it's not making much sense.
(full disclosure, I’m a techie who’s gradually woken up to the idea that the tech might just be the most abused way to exploit people)
It's a bit ironic, because a lot of tech offers partial compensation in stock. Something else that really doesn't happen in games unless you work for like, the 3-4 largest studios. So they should at least understand that your compensation is not all based on labor for time worked.
that's gone out the door in the digital age. Compaies at this point have spent centuries trying to enfoce this model while witholding stuff like stock and royalties to take a part of what the company enjoys by protifting for decades off of a single (underpaid) piece of labor.
I don't exactly sympathize with a robot now trying to do the same. Pay your labor.
I don't know. Anyone funding the work is accepting a risk.
> Should they not have been paid after 1991?
They definitely should get paid for their shows and live performances. The band itself can't be copied. Artists are extremely scarce.
Their art, however, is not. Once created, the scarcity of their recordings is artificial and fundamentally time limited anyway. Even if I were extremely tolerant of copyright, I'd argue for a term of only 5-10 years maximum with absolutely no possibility of extension.
In other words, even if we accept copyright as legitimate, they sure as hell shouldn't still be getting paid for some late 80s album. They've already been adequately compensated for those creations. If they want more, they should have to keep making new stuff so that they can benefit from new copyrights which will also expire after a short time.
Creators are not supposed to be able to strike gold once and then enjoy eternal royalties. Copyright must have short time frames or it's in breach of the social contract. The reality is we're doing creators a favor by pretending that it's hard to copy their stuff so they can make some money. We do this because they assured us that eventually all of it would belong to us: works would the public domain.
The copyright industry isn't keeping up their end of the bargain. They continuously pull the rug out from under us by extending copyright to the point we'll be long dead before our culture is returned to us. It's offensive and we should all stop pretending. They need reminding that public domain is the natural and default state of all intellectual work.
> How would you predict that value before its creation (or even after)?
I'd look at the artist's past work. If there is no past work, then I don't know.
> If you're saying that only the labor has value
I'm not saying that at all. Creations are valuable. Creators are valuable. The labor of creation is valuable.
Value is assigned to stuff by humans. Obviously humans value art. The price however is given by supply and demand. The fact is that supply of intellectual works approach infinity after they are created and therefore their prices approach zero. So it makes perfect sense to assign prices to the labor of creation but zero sense to assign a price to the product of creation. Copyright is an exercise in denying reality.
> and all labor is valued equally
I definitely did not say that. All labor is different. I value some creators a lot more than others. Some creators I don't value at all.
> that sounds sort of like marxism
I must apologize if I gave that impression. I hate marxism.
Why not? The fact is that even if the album is free, there will be people paying spotify $10/month to listen to it on demand. How is it fair that Spotify can profit from it for decades to come because they offer convenience, over the artist who made the music 10 years earlier and now relinquishes their art not even a quarter into a typical career?
Copyright is abusrd now, but it's not a bad concept. I think the original copyright law of 14 + 14 worked well enough. Life expectancy increased so I'd increase it to 14 + 14 + 14 (or 10 years after the death of the original author, whatever comes first). You fund an artist for their typical career length (if they choose to extend twice) and once they are (near) retired the song is free to work off of. In the meantime you simply negotiate if you want to use their work.
Copyright means that you need to at least pay that artist you stole from in some way, which the government enforces so artists don't stop creating.
CliffNotes, Wikipedia, etc. have huge quantities of summarized copyrighted work.
https://en.wikipedia.org/wiki/Derivative_work
Second, you ignored the "purely derivative" bit. You have to look at to what extent the use is derivative or transformative. See https://en.wikipedia.org/wiki/Transformative_use for a bit about that. (Note, this is a legal term defined by various precedents. OpenAI can't just argue, "Turning it into an LLM is a transform, so it is transformative!") Since CliffNotes is educational and Wikipedia is nonprofit, it is relatively easy for both to qualify as transformative.
As a result your response underscores the point that was made. There are a lot of shades of grey. You really can't just seize on a couple of phrases and key points, then jump straight to the answer. You have to understand how the courts will decide, and then accept that there is an actual judgment call whose outcome depends on the judge judging.
(I'm not a lawyer, but I have had excessive exposure to them in the past.)
Is there data that supports this? I’d be interested to know what % of people who buy a Cliffs Notes have already _bought_ the original.
But, anecdotally, it's what I've seen to be the case.
Yes.
For example, Wikipedia cites many research journals that otherwise are available only by subscription.
Prior to Wikipedia, gated information centers were the norm.
The question was NOT whether it spreads information from the articles to people who wouldn't have paid for it. The question was whether it suppresses sales of the articles to people who otherwise might have paid for it.
That's a more complicated question of fact. Some people now read Wikipedia and won't buy the article. Some people encounter the reference on Wikipedia and decide to buy the article. Which happens more?
I don't have data. But publishers do. And https://scholarlykitchen.sspnet.org/2022/11/01/guest-post-wi... shows what publishers concluded.
Publishers concluded that Wikipedia references are good for sales. And so jumped on the chance to cooperate with https://wikipedialibrary.wmflabs.org/. Which is therefore able to give free access to 90% of subscription only databases to you if you can prove that you're the kind of person who is likely to add citations to Wikipedia.
Legal questions are funny like that. You have to answer the question actually asked. If you merely answer another one that sounds similar to you, your answer is generally wrong.
You're the one presenting unfounded claims with confidence here. There is well established case law about not being able to copyright facts. If you are actually fully paraphrasing a presentation of facts / ideas and not just altering a couple of words here and there, then there is a very strong case for non-infringement.
No, I'm not. On the contrary, I'm really looking forward to this case because I believe it will be a great test of a bunch of concepts that are totally novel in the world of copyright law as it applies to generative AI. The only things I am presenting with confidence are:
1. That anyone who declares that something is unambiguously fair use (or, contrarily, unambiguously infringing) is likely wrong. There is simply too much latitude by judges, and there have certainly been cases where a ruling went one way, only to be overturned on appeal.
2. While I certainly have an opinion on how I think this case will be decided, I'm not presenting that with unwarranted confidence. Instead, I linked that great article on the 4 factors of fair use determination because it's clear to me lots of people are saying "fair use!" on one side or the other with no understanding of the factors judges must actually consider when making a determination.
However, none of that matters in this particular thread. There are well established precedents about paraphrasing news articles and they do not support the claim you made
Remember. The NY Times does not have a record of filing frivolous lawsuits. Particularly not against companies with deep pockets. So it is almost certainly true that a lawyer who knows the law better than you thinks that this has a real chance. So you should be looking for flaws in trivial defenses that you can think up, rather than assuming that you know best.
For example take your copyright facts defense. That would be great if the NY Times was a phone book. They aren't, in addition to facts they offer analysis, editorial positions, and so on. For example I just asked ChatGPT, "In 2016, did the New York Times generally support or oppose President Trump?" I got back an answer talking about various kinds of concerns that the New York Times had, including an editorial titled, "Why Donald Trump Should Not Be President". The copy that ChatGPT needed to have to do that has a lot more than just facts in it.
Now if you paraphrased the NY Times like ChatGPT did when it answered me, you'd have a perfect fair use defense. But you aren't doing it for money, you didn't make a copy of all the NY Times, you aren't destroying the market for the NY Times, and you're legally able to own copyright in your transformed work. OpenAI is doing it for money, did copy all of the NY Times, is seriously impacting the market for NY Times articles, and ChatGPT generated text does not get a copyright.
Fair use is filled with shades of grey. Even if ChatGPT appears to do the same thing that you do, it is far less clear that OpenAI will enjoy the same level of fair use defense.
> They aren't, in addition to facts they offer analysis, editorial positions, and so on.
Those opinions and ideas are also not copyrightable. Only expressions of them are copyrightable, which is why paraphrasing facts, ideas and opinions is not a violation of copyright.
> Fair use is filled with shades of grey.
Yes, but not all those shade are equal. There is a long history of litigation showing that paraphrasing news articles is fine.
This is the weakest part of the case(s) against OpenAI. "Derivative work" is a legal term of art meaning a direct adaptation, like writing a screenplay of a book or translating a book into another language.
NYT has a stronger case than Sarah Silverman here because they can show actual 'memorized' text rather than just summarization, but given that those memorizations are a) an unintended failure mode of the training process, and b) from an older version of the model that has been updated to no longer regurgitate memorized text, it's not really clear how in current form GPT could possibly be considered a derivative work.
The latter is more defensible.
On the other hand, it's understandable why NYT is worried. OpenAI itself says that occupations like: Writers and Authors, Web and Digital Interface Designers, News Analysts, Reporters, and Journalists, Proofreaders and Copy Markers are "90-100% exposed" to what OpenAI is building.
https://www.nytimes.com
https://www.cnn.com
https://www.washingtonpost.com
https://www.reuters.com
https://apnews.com
I don't care that the car replaced the horse carriage because it didn't need to compensate horses nor handlers to do so. AI being the newest iteration of scraping data from artists, writers, etc. to profit millions off of is directly using the "horse handler's" work. If these LLM's threw NYT a royalty to use their articles as training material, there wouldn't be a lawsuit.
I have worked on many documentaries and any time we said “fair use” internally what we were implicitly saying is “nobody will come after us because they know that we are probably safe under fair use if this escalated.“ But again, we could never preemptively apply it. We were just anticipating potential conflict and gauging how likely it was to occur.
In the US, whether or not you make money has little to do with whether or not your use qualifies as "fair use".
That a use is noncommercial is often a deciding factor in the success of a fair use defense. GP is overstating it though, since it’s still one of many factors.
But what I was arguing was that a use is not "fair use" merely because it's noncommercial in nature. I cannot make copies of movies and give them away on the street for free and successfully claim "fair use".
Weird Al has made a fantastic living copying music while only changing lyrics. He makes very heavy use of the satire plank of Fair Use.
The “commercial” test is only part of the decision criteria for Fair Use.
The "commercial" test is only a part if the criteria and not necessary, but to say it has little impact is clearly false.
"Does Al get permission to do his parodies?
Al does get permission from the original writers of the songs that he parodies. While the law supports his ability to parody without permission, he feels it’s important to maintain the relationships that he’s built with artists and writers over the years. Plus, Al wants to make sure that he gets his songwriter credit (as writer of new lyrics) as well as his rightful share of the royalties."
The fact that he could rely on fair use is separate from whether he as an artist does rely on fair use.
Not only does he get permission from the original authors, he also pays royalties to the authors despite legally not having to do so.
If that were true, I could take a band that I hate, copy all of their music note-for-note, then release an exact copy on the market and undercut them by selling their entire discography for $0.01
Fair Use requires one of several enumerated activities, including satire, education, journalism. You can’t just copy content and hope that it passes Fair Use.
Hire a lawyer if you are unsure. But at least read the Wikipedia article on the subject if you are going to talk about it.
Based upon what? You think other publishers use NYTimes articles for free without license?
Presumably, if it can remember at least a paragraph or two of each article, then surely the same would be true of any text it ingested and the model size would approach the dataset size (probably actually much larger). I don't believe this is the case at all, even searching around, I've not found any good recent examples of it regurgitating copyrighted text verbatim.
It's cool to hate AI stuff if you're a creative atm. But gotta love those generative/algorithm based PS brushes, that's still real art!
"Indeed, the opening paragraph of "A Game of Thrones" by George R.R. Martin, with the chapter titled "Bran," starts as follows:
"The morning had dawned clear and cold, with a crispness that hinted"
And then it cuts off, whether that's because OAI now have an oh shit filter or just the model had access to the first page or publicly available articles quoting the first line, I'm not sure.
I tried other chapters and random sections and it could get a sentence or two right but then hallucinated; what's more likely NYT and GRRM? That your works are being reproduced verbatim? Or that Facebook, YouTube descriptions, fan tumblrs and hell, the publicly available and multiple GoT related wikis that include a variety of passages from the books were used as training data?
"It could be fair use if conditions a, b, and c are met. Condition a means..." ;)
I think what wouldn't be covered is reproducing substantial portions of an article, especially if it's done without attribution. Tier 2 publications that fully reprint NYT or AP/Reuters articles are usually doing this via a paid News Service or Content License. See: https://nytlicensing.com/content/new-york-times-news-service...
I hope the NYT prevails here, personally. Models will (and are) currently tainted by data they should not contain and for longer term privacy concerns this needs to be addressed early and have significant consequences or we're headed towards a world where this type of technology will make our ad-targeted world seem like a much more manageable past.
You just described Google. When you think about it, it's surprising that Google is legal. However, it is well established that what Google does is perfectly legal. Remember that internally Google keeps and uses complete verbatim copies of every web page they index.
Yes, Google offers a link to the source. If OpenAI did the same, even if only 0.1% of people clicked on the links and NYTimes hardly got any revenue from it, would that make it legal in your eyes? What if they implemented a system that detected when it was outputting a verbatim copy of something and simply paraphrased it? NYTimes clearly doesn't have copyright on paraphrased versions of their articles. I think it would be pretty silly if the government forced them to do that as it wouldn't make any practical difference to anyone.
Google has a wide range of products and shakedowns. Not all of them are "perfectly" legal: Google is being challenged in court over some of their shakedowns and products practices.
Paraphrasing is also known as cloning and is often a copyright violation
In US copyright law facts cannot be copyrighted, so copyright on factual content like newspaper articles is limited. Simply replacing a few words wouldn't work, but I am certain that GPT-4 is capable of paraphrasing factual content at a level that would not be considered infringement if a human did it.
Genuinely - what are you talking about besides your own assumptions? you just assume everything google does is legal and therefore any one else doing anything arguably similar must also be legal? Without regard for factual details that do matter to copyright law? Such as license?? Your own description of copyright law here is very stunted - you can't paraphrase articles of the NYTimes and call it a fair use. You can report on what the NYtimes reports on... because that's what news is.
Not an assumption. This is well established. They've been doing it for twenty years!
> Without regard for factual details that do matter to copyright law? Such as license??
What license? Google doesn't in general have or need an explicit license to crawl websites and neither does OpenAI.
It's not at all well-established. How many anti-trust suits is Google facing now? Your proposition defies common sense.
>What license? Google doesn't in general have or need an explicit license to crawl websites and neither does OpenAI.
It's not the crawling the website that OpenAI did that it needs a license for... why bother conversing if you are going to be this obtuse?
Seems like the legal answer is unclear but, like Napster, such a system seems like it would lose in court.
One site clones fox news. One clones news max. And so on, cloning many news sites, sports sites, any news site. Automated, massive scale content farming. Think of the websites recommended by Taboola but, realistically, a whole lot worse.
Nobody is seriously going to ChatGPT and trying to trick it into regurgitating old NYT articles as an alternative to paying for access to NYT's archives. Meanwhile, newspapers went as far as getting the laws changed in several countries because they felt Google was competing with them too much and didn't like the fact that it was legal.
Can they? Here's reference to a legal fight where Google scraped song lyrics from a lyrics website, and presented the lyrics verbatim directly to users (bypassing the original site and the ads that allowed that site to operate)
https://www.rollingstone.com/music/music-features/genius-law...
The whole point of a search engine (as we've classically known them) is to index the web and respond to queries with a list of links that you will inspect and click through on. The whole point of an LLM chatbot tool is to eliminate those inspecting and clicking-through steps, becoming a one-stop shop for content whose substance was created by someone else. That's also the whole point of GP's hypothetical, which is why it works as an analogy.
---
There are substantially better arguments for search engines being legitimate fair use. Consider, for example, transformation. AI defenders will argue that these systems are transformative because they reshuffle elements of their input in their output, but that's clearly a much weaker form of transformation than one in which the transformed work has an entirely different nature and purpose, i.e. search engines vs. the results they return. Ultimately these technicality-based "nuh uh" arguments aren't going to save the practice of training AI on unlicensed data, because they are incompatible with the spirit of copyright law even if the novel nature of these technologies means the letter of said law can't quite nail them down yet.
If these arguments do succeed, it will be because the judicial/regulatory environment in which they were applied has been corrupted by capital.
An LLM takes an input string a corpus of text, and returns a series of text that best comes after the next input string.
To get a paragraph of output, you run the search over and over again
Both the search and LLM reshuffle the inputs to the outputs.
If I'm describing the purpose of the LLM, it's got a wide number of usages. "Making my resume look more professional" or "be a crud api" or "reformat my ask into a api call to X service" or "give me a timeline of events surrounding Y with source links"
If I'm describing the purpose of an 18 wheeler, it's got a wide number of usages. "Carry my chicken" or "carry my lettuce" or "carry my Cheetos". Or, simply, "carry my groceries".
And how did the training data contribute to the content in any meaningful way? Inspiration isn't substance.
You think all fantasy writers gotta pay Tolkien estate bc so much of fantasy draws from his tropes? Lmao no.
If training data is so unimportant, why not simply not use it and avoid the controversy? At the very least that would certainly fix the issue where the model demonstrates how "inspired" it is by NYT articles by reproducing them verbatim.
:)
Possession is not a crime when it comes to copyright. It's not like physical things (e.g. drugs or guns) at all. This is why comparing copyright violations to theft is silly.
ChatGPT can absolutely keep verbatim copies of the entire works of basically anything and not run afoul of the law. When it regurgitates a small part of an article that's covered by fair use in theory but the truth is that fair use can only be determined by a judge in a court of law when someone is sued. It cannot be determined with any sort of certainty ahead of time. It's a legal defense, nothing more.
Summarizing content has been legal forever as well (see the other posts here talking about Cliff Notes and some similar products). That's not even fair use that's just like, people's opinions, man (legally speaking).
I don't think the NYT will get what they want out of this at all.
Man thinks piracy is legal
That's not a good question.
If I look out of my window and see my neighbor go to the shop, that's fine. If I use cameras and track everybody I see on the street and put them in a database, then that's problematic and illegal in many places.
Logic does not necessarily apply when scaling is involved.
But what I'm saying is that answering the question does not allow you to deduce anything about your rights; that's what I mean by "not a good question".
If we want to establish whether scenario A is fair use or not, and we all agree that A is "worse" (regarding fair use status) than some other scenario B, then if we also agree that B is not fair use, A by definition isn't either. The opposite is not true, of course: B being fair use does not imply that A has to be as well.
I find that kind of upper/lower bound logic can be pretty useful and I think it's what the parent comment was trying to do.
On a related note, that same logic is why I think Godwin's law can be a bit misapplied now and then. Sometimes bringing up nazis/Hitler can be useful to establish some ground truth in a debate (instead of just a way to imply your opponent is actually a bad person, or, possibly, an actual nazi themselves). E.g. a conversation on the morality of violence is vastly different depending on whether you agree that violence against nazis is ok or not.
Afaik not illegal in the US. You put a camera on your own private property (window), use it to record what’s happening in a public space (the street outside), and then store that data in a database (that other people can presumably access). Unless I am missing something, this scenario is perfectly legal in the US.
Inbefore I get hit with “not every country is like that at all,” NYT is based in the US and the lawsuit is filed in the US. So how a bunch of other countries deal with similar issues shouldn’t really have as much bearing on this specific case.
So like....Wikipedia, CliffNotes, encyclopedias, etc?
None of these pay royalties to original.
> Implications: The Ninth Circuit's declaration that selectively banning potential competitors from accessing and using data that is publicly available can be considered unfair competition under California law may have large implication for antitrust law. [citation needed]
> Other countries with laws to prevent monopolistic practices or anti-trust laws may also see similar disputes and prospectively judgements hailing commercial use of publicly accessible information. While there is global precedence by virtue of large companies such as Thomson Reuters, Bloomberg or Google [or LexisNexis or Westlaw] effectively using web-scraping or crawling to aggregate information from disparate sources across the web, fundamentally the judgement by Ninth Circuit fortifies the lack of enforceability of browse-wrap agreements over conduct of trade using publicly available information.
NYT articles are largely behind a paywall for everyone. That means they are not publicly available, and a competitor who was blocked from accessing or reproducing that content without a license would not be "selectively banned"
Consider the analogy from libraries that want to do data mining.
"Unfortunately, in licenses for digital scholarly content the majority of content acquired by research libraries publishers often include terms that prohibit certain uses that would otherwise be allowable under the Copyright Act. For instance, licenses may require libraries or individual researchers to negotiate for otherwise lawful activities, such as text and data mining, and to pay exorbitant fees on top of the cost of the content itself. While new regulations allow researchers to circumvent technological protection measures to access copyrighted materials, licenses for that content may include terms that explicitly prohibit this circumvention. In many cases, these activities might actually increase the value of published material; for instance, if a data-mining project yields new knowledge about a topic covered in a journal, it may very well spark new interest in that journals content. Libraries and publishers have often assumed that license terms that restrict copyright exceptions are enforceable under state contract law. There is, however, surprisingly little case law on this point."
https://www.arl.org/wp-content/uploads/2022/07/Copyright-and...
Putting some string in a robots.txt to try to stop data collection is an amusing "solution". Should copyright owners have "Terms of Use" that limit usage for commercial "AI" purposes.
Why should the law treat a LLM in a body reading NYT on a tablet differently than a LLM browsing the content from a website online and reading that?
While harder to do as a human, if memorised a copyrighted book and then did a live reading on TV, or produced replicas from memory and sold them (the most comparable example), I’d be sued.
Humans produce derivative work all the time, and it’s fine for LLM’s to do that, but you can’t do it verbatim.
This is not the most comparable example, because it's not what ChatGPT is doing. The most comparable example is if you were hired as a contractor and the employer asked you to write verbatim some copyright content you'd memorised. If the employer then published it, they'd be the one liable, not you.
>Humans produce derivative work all the time, and it’s fine for LLM’s to do that, but you can’t do it verbatim.
Nobody's suggesting preventing humans from consuming any copyrighted content just because in future they might recite some of it verbatim, but that's what NYT want for LLMs.
No, you'd both be liable. You are not allowed to create copies of a copyrighted work, even from memory, for any commercial purpose. Making it public or not is irrelevant.
This is more obvious with spftware: if I copy a version of AutoCAD that my previous employer bought and sell it to another company, or even just use it for my current employer without showing it to anyone else, I am violating the copyright on that software, and I am liable. Even though obviously no "publishing" happened.
Similarly, if you hire a decorator to paint Mickey Mouse on the inside walls of your private kindergarten, the decorator is violating Disney's copyright just as much as you are, even if neither of you has made that public.
That's the point at which infringement occurs in your example. It's not the memorizing that's the infringement, it's the reproduction from your memory.
We shouldn't be regulating your hippocampus encoding the book, but your reproducing the book from that encoding.
Similarly, we shouldn't be regulating the encoding of material into the NN, but the NN spitting back out the material.
Are they all owned by one mega-corporation, which is going to do as capitalism does, and use them to squeeze money out of all of us? Then I'm happy to ban them.
The opportunity cost of holding this technology back is going to literally be millions of people's lives given current trends in its emerging applications.
Police usage, not training.
You'd get the same problem with someone with a photographic memory who a group of people would turn to recite them the news instead of buying the newspaper.
As of now public performance of copyrighted material is infringement.
I fully agree with the perspective that infringement in usage needs to be limited even if I strongly disagree that training is infringement.
My guess is that the court will likely find in the Times favor, because the legal system won't be able to understand how training works and because people are "scared" of AI. To me, reading a book, putting it in some storage system, and then recalling it to form future thoughts is fair use. It's what we all do all the time, and I think that's exactly what training is. I might say something like "I, for one, welcome our new LLM overlords". Am I infringing the copyright of The Simpsons? No.
I am guessing some technicality like a terms-of-use violation of the website (avoidable if you go to the library and type in back issues of the Times), or storing the text between training sessions is what will do OpenAI in here. The legal system has never been particularly comfortable with how computers work; for example, the only reason EULAs work is because you "copy" software when your OS reads the program off of disk into memory (and from memory into cache, and from cache into registers). That would be copyright infringement according to courts, so you have to agree to a license to get that permission.
I think the precedent on copyright law is way off base, granting too much power to authors and too little to user. But because it's so favorable towards "rightsholders", I expect the Times to prevail here.
Anyway, like I said, I don't think OpenAI will win this. Someone will produce one verbatim article and the court will make OpenAI pay a bunch of money as though every article could be reproduced verbatim, and AI in the US will be set back that many billion dollars. It probably doesn't matter in the long run; it preserves the status quo for as long as the judge is judging and the newspaper exec is newspaper exec-ing. That's all they need. The next generation will have to figure out how to deal with AI-induced job loss... and climate change. Have fun, next generation!
"Its what we do all the time" is a major assumption
So, if you setup a service like ChatGPT but powered by humans responding real time to queries, and these humans would occasionally reproduce large chunks of NYT articles, they and the service itself would be liable for copyright infringement. Even if they were all reproducing these from memory.
Now, this is somewhat different from the discussion of whether training the model on the copyrighted data, even if it had effective protections from returning copies of it, constitutes copyright infringement in itself. I believe this is a somewhat novel legal question and I can think of no direct corollaries.
I certainly don't think we can just handwave and say "at some level, when a human reads a copyrighted work, they are doing the same thing", because we really don't know if that is true. Artifical neural networks certainly have no direct similarity with the neural networks in the brain as far as we can tell. And, even if they did, there is no reason to give a machine the same rights that a human has - certainly not until that machine can prove sentience.
To put it another way, let's say I turn the dial all the way the other way, I train the worlds crappest LLM on NYT material, it massively massively overfits and all it will ever return is verbatim snippets of the NYT. Is that copyright infringement?
The core part of the argument here is actually just that OpenAI doesn't want to adhere to what the current standard is for using copyrighted material, if you want to use it and create something new with it you need to license the material. Since OpenAI's LLM isn't actually like a human it needs to license such a vast dataset that it would be uneconomical to run the business without stealing all the content.
Sure, when something is clearly derived, or just expressed in a new medium, then I'm sure it's still covered. But if it goes through an LLM and the result bears little resemblance, how can that still fall under copyright?