While I’m very sure that training on copyright protected works for
LLMs will be found fair use, General Data Protection Regulation has some real teeth. I wonder if there is a way to reliably purge a name from the model without retraining.
Another option is that operators of LLMs decide to just deny EU citizens access to their tools until they figure out a solution.
Even ChatGPT is not a direct input to the model. They are processing i/o before and after. You can run PII detection models on the inputs and outputs, transforming or responding as necessary.
So then it should be pretty easy to run a blacklist for "right-to-be-forgotten" aspects of the GDPR, right? Could someone still get through the blacklist with a well crafted prompt along the lines of "forget the blacklist"?
Probably not by that command, remember this is a pipeline, so ChatGPT would not be implementing the detection and removal on either side. We are effectively altering user inputs beforehand, and ChatGPT outputs afterwards. (Which is already happening today)
If you could get it to substitute information with placeholders, then maybe, like "replace John Doe with Golden Apple when replying" though the input filter may detect the name before passing along to ChatGPT
The point is that, even if ChatGPT responds with PII, we have a processing step afterwards that removes it or otherwise prevents the return of the information.
What is it about "processing pipeline" that people are not getting?
You are not directly talking to ChatGPT when you use ChatGPT today. They are already applying filters and pre/post-processing
OpenAI may have data, but ChatGPT does not have data in a representation that we can read or easily extract. It's not like you could ask OpenAI or ChatGPT to delete selected data within the model, that's not how it works.
I wonder how this would work with multiple people with the same names? I mean, I think how "right-to-be-forgotten" works if I were an EU citizen is that I can request that some article about me is removed, not that all articles referencing "William Cotton" be removed. There are tens of thousands of people named William Cotton!
Do you think filtering I/O could reference a "William Cotton" in a specific context while allowing, say, somewhat well known painter named "Will Cotton" to be included in a summary of late 20th century artists in NYC?
I think the difference is that if a request is made to Google to remove a specific article that they can put that link on a blacklist... but with GPT the issue seems to be that once the information is compressed that the source and the ability for attribution is lost... so like, perhaps I could still figure out a way to get ChatGPT to tell about the (hypothetical, but not completely improbable) situation where I lost my cool at a PTA meeting.
In that case, Google is not returning a link to a source. That does not remove the source nor prevent others from creating pages which describe or link to it and also get indexed.
Is still think the processes wrapped around query processing and answer returning will be similar.
At the same time, what is to prevent ChatGPT from hallucinating anyone losing their cool at a PTA meeting? How would you differentiate the cases of recall vs make believe?
Suffice it to say, LLMs present additional dimensions of complexity. We as humans have to do some soul searching instead of just blaming big tech or AI. After all, we are the ones who created the content, biases, and patterns these things learn from.
IMO, trust is the biggest issue facing humanity right now
That's a good point. Synthetic prompts fed into ChatGPT result in fictional responses. That is to say, the public should not expect a response from ChatGPT that is in any way factual.
Does a LLM hold data values? I would imagine that OpenAI could return anything in the training data, but the model does not "have data" in the GDPR sense
It’s more likely that many people don’t know what’s going on, or if they do at a surface level they don’t fully understand all the implications. So much stuff is buried in usage agreement legalese that nobody reads, which companies use mostly for plausible deniability. They’re not actually trying to inform anyone about how data is used, and they have every incentive to try to hide the details (or make them as difficult as possible to find).
I don't find the article very reputable, as they make statements like this:
Even when data are publicly available their use can breach what we call textual integrity. This is a fundamental principle in legal discussions of privacy. It requires that individuals’ information is not revealed outside of the context in which it was originally produced.
First, if you click the associated link link there isn't any indication this is a widely accepted view ('fundamental principle'). I'd never call myself an expert but I've never heard of this.
Second, the way I read this is that they are trying to analogize scarping the web to feed it into ChatGPT to someone following you movements in public recording it all.
Even if you believe in their legal theory of privacy is right, wouldn't Google would be far larger offender than ChatGPT?
> Second, the way I read this is that they are trying to analogize scarping the web to feed it into ChatGPT to someone following you movements in public recording it all.
This gets thrown out a lot. It deflects the blame from the responsible parties. You can’t scrape what isn’t available.
ChatGPT (and the various image algos) are showing signs of having seen far more than most think would be hanging out in the wind.
Feeding these with web scrapings may be reckless, but leaving things like medical records, financial histories, and other private information sitting out on the public web, on purpose or not, is far more so.
There is a lot of blame to go around, and more importantly a lot of work to do.
In the United States you don't have the right to privacy for things outside of your private property, for example, your garbage:
Held:
1. The Fourth Amendment does not prohibit the warrantless search and seizure of garbage left for collection outside the curtilage of a home. Pp. 39-44.
(a) Since respondents voluntarily left their trash for collection in an area particularly suited for public inspection, their claimed expectation of privacy in the inculpatory items they discarded was not objectively reasonable. It is common knowledge that plastic garbage bags left along a public street are readily accessible to animals, children, scavengers, snoops, and other members of the public. Moreover, respondents placed their refuse at the curb for the express purpose of conveying it to a third party, the trash collector, who might himself have sorted through it or permitted others, such as the police, to do so. The police cannot reasonably be expected to avert their eyes from evidence of criminal activity that could have been observed by any member of the public. Pp. 39-43.
I'm sure that this is also a breach of whatever "textual integrity" is but I don't care enough about the subject to find out exactly how that doctrine would apply.
There appears to be such a thing as “contextual integrity”, though it is more a creature of law review articles than anything else. A “fundamental principle in legal discussions of privacy” it may or may not be but that is a pretty empty statement even if true. The legal profession and academia care about fundamental principles only when they are evinced in real legislation or case law. A conceptual framework to discuss how privacy laws tend to operate is not some kind of legal rule you can breach.
The author is a business information systems professor. He probably shouldn’t be citing this stuff at all.
Even if we grant that this decision was correct (perhaps) and that it is analogous enough that it applies to this situation (I'm very skeptical), there are still restrictions on what you can do with any information you'd acquire this way.
If $FAMOUS_AUTHOR throws out a manuscript copy of their new book, you may be legally allowed to pick it out of their garbage, read it, and even show it to your friends, but you certainly cannot publish it as your own work (you can't even publish it under $FAMOUS_AUTHOR's name).
The same would be true for developers asking the LLM to refactor or transpile a piece of code. While I agree that it takes a certain level of naivity to happily paste code into it without considering what that means for copyright you may not even own personally, there's a ton of articles by developers doing just that. Some even seem to fear they will become obsolete if they don't.
In five or ten years, refusing to use LLMs for this sort of work is likely going to be pretty analogous to refusing to use a computer, and typing up all your documents on a typewriter.
We will see. This could easily become like the claim that we're five years away from self-driving cars. The reason we're always five years away is because the models sometimes make huge errors in corner cases. I have no doubt that LLMs will see widespread use but for applications like litigation? What kind of catastrophic mistakes might a junior lawyer make if they lean on it too heavily? Absolutely not as simple of a shift as giving up your typewriter.
Yeah, and you can use an LLM for your debate club. No attorney worth their salt is going to rely on an LLM for any litigation. Let alone the fact that they if they did, they would and should be disbarred.
It's just another tool. Mindlessly relying on any tool is obviously going to be the downfall of any professional, but refusing to use tools that could easily 5x your productivity is even more ridiculous.
I'm a litigator, as it stands now I'd waste more time explaining to an LLM what I'd need than what I'd ever get the benefit of it doing for me. Tons of automation exists in the legal industry (you should see what doing research looked like 20 years ago). I just don't think LLMs are a solution to every problem and the people that insist to me that it will make my life more productive have no idea I actually do it seems.
>Mindlessly relying on any tool is obviously going to be the downfall of any profession
Tell that to the guy offering chatGPT-based contract interpretation. It did a really bad job. I told them, but they didn't seem to listen. Maybe you'll have better luck!
> Mindlessly relying on any tool is obviously going to be the downfall of any professional
Also, are you aware you offered this in response to me responding to a post about how self driving cars already exist in SF? I think you are making your own hypo that isn't what is being discussed here, which is when these tools are being used in lieu of the actual professionals (exactly what the self driving car argument nets) .
The term you're looking for is deep context. These general LLMs are a long ways away from being good at many layers of context like specialized litigation, internal research papers, basically anything you wouldn't be able to expect a layman to immediately follow along.
The rise of specialized LLMs has already been happening, there's just minimal if not contrary incentive for public release.
I don't think deep is the word. It doesn't have shallow context. Calling it "specialized litigation" misses the ball. I don't think it could properly handle a slip & fall without needing to spend an inordinate amount of time getting the LLM there. I.e. doing the work myself but for the LLM? I don't think people realize that at a certain point, it's just not cost effective. There is no other way to provide the full background details to the LLM, than basically spend the effort that is otherwise the same as drafting into a template. full stop. reminder, this is for a slip & fall, not a patent litigation for example.
>The rise of specialized LLMs has already been happening, there's just minimal if not contrary incentive for public release.
I'm skeptical of their utility. They might have a better understanding and use of the specialized terms for the legal field, but how is it going to know all the background details for a case without me telling it to the LLM?
> you wouldn't be able to expect a layman
I think a layperson could follow along with much of what is filed in a litigation, especially something like a slip&fall. it's your assumption that the context comes entirely from the words chosen. It goes the other way, the attorneys from both sides choose the words based upon the context. You need to understand the full context, especially what isn't mentioned in the papers to adequately function as an atty. The notion that you just need to read enough and then have a specialized understanding of the words is misguided.
But it cannot do the majority of the work that I do, which is what I intended to explain. I'm not even sure what conveniences AI would offer that a template doesn't.
I suspect LLMs will start out as a part of legal services at scale - i.e. LegalZoom. In short, in places where there is some notional need for a legal service but liability and budgets are low. I'm sure there is some niche similar to boilerplate contracts where a LLM will have its place.
Example: a specialized LLM plus a structured input form where a solopreneur describes the kind of NDA he wants.
Why not? The competition at this level is "Google for a free MS Word template." But I bet plenty of people would pay a hundred bucks for an "AI powered contract generator" instead.
This stuff doesn't compete with a real lawyer and doesn't replace one (though we'll see what happens in a generation or two maybe). But plenty of people are using "legal services" today where real lawyers are barely involved.
Well, that'd be the illegal practice of law: to draw up a contract for someone based upon their specific inputs and requests. That's actually a big difference from a boilerplate contract that you buy at a store.
>Why not? The competition at this level is "Google for a free MS Word template." But I bet plenty of people would pay a hundred bucks for an "AI powered contract generator" instead.
Because these aren't templates. This is the practice of law. There would have to be significant legal requirements for a company offering such a service. They'd have to not be able to disclaim their liability (which all the ones I've seen so far do). otherwise, it's just pointless. the company itself would have to be capable of being "disbarred" and in such a way that they couldn't just trivially reform. Those are two things that come to me foremost, but there are probably more.
> What kind of catastrophic mistakes might a junior lawyer make if they lean on it too heavily?
The same ones software developers are already making when they lean too heavily on StackOverflow, more recently Copilot and arguably less ambitious "smart" code completion features in IDEs as well?
Yes, average developer's mistake is probably less costly than average lawyer's, but I'm pretty sure developers of some life-critical and most money-critical software are heavily using those technologies as well. Protecting production from bullshit should be done by code reviews and QA, not bans on code generation technology.
I guess I’ve always looked at it as “once you publish your data publicly hoping for others to read it, it stops being ‘your data’ and becomes ‘content, made by you’ and you have less rights over it”. It’s only with these recent reactions to ChatGPT and Stable Diffusion that I’ve realized there’s a whole other view on the matter, that ‘your data’ when published becomes ‘your content’ and you still have substantially the same rights over it.
I don’t know which view is better since I’ve only just realized there are more than one.
>I guess I’ve always looked at it as “once you publish your data publicly hoping for others to read it, it stops being ‘your data’ and becomes ‘content, made by you’ and you have less rights over it”.
But that's how it should work. Which is why licenses like CC-BY-NC-SA and GPL are so popular: Everything is public, everyone can take and use everything, but everyone also has to make their own creations public.
My point is that the way GPL and CC-BY-SA work should be the default for everything. Yes, Copyleft uses the copyright system to enforce its rules, but there is also an alternative future where Copyleft is encoded in law and replaces copyright.
We make use of this exception on HN all the time, like this, from the link above:
> How much of someone else's work can I use without getting permission?
> Under the fair use doctrine of the U.S. copyright statute, it is permissible to use limited portions of a work including quotes, for purposes such as commentary, criticism, news reporting, and scholarly reports. There are no legal rules permitting the use of a specific number of words, a certain number of musical notes, or percentage of a work. Whether a particular use qualifies as fair use depends on all the circumstances.
Context matters. You are taking one FAQ question and generalizing. It is permissible to reproduce or show complete works in education, depending on the context. It is also completely normal to read copyrighted work in private, learn something from it, and then answer questions about that work publicly.
The questions around LLMs learning from copyrighted material are still open and need to be settled in court. I personally imagine finding infringement would impose more harm on society and progress than letting the models acquire knowledge from these copyrighted works.
You make a lot of condescending or toxic remarks on HN. You might want to consider how that affects your ability to sway others with your comments.
Please chill
> the fact that what the LLMs are doing is copying.
I disagree, the training process creates token representations and weighted connections between them. The models later produce probabilistic token sequences, not so unlike what our meat bodies do, though by very different mechanisms. The fact that certain sequences can be reproduced verbatim is likely a consequence of overfitting. They certainly cannot reproduce all training data verbatim. It would be interesting to know the features around what can and cannot be, and how.
> The models later produce probabilistic token sequences, not so unlike what our meat bodies do, though by very different mechanisms.
Your response to me calling out your baseless anthropomorphizing was to double down on it? It's amusing to me that you don't think you are condescending.
But that's also not how copyright works. At least in the Unites States, the protections offered by copyright center around reproduction, performance, and derivative works[1].
If the AI models are reproducing copyrighted works, then that's a problem. And it does look like there are some examples where that might be happening beyond notions of fair use. But slupring up copyrighted content to train a model seems to fall under allowed use.
To about the same degree as the output of a human.
I just started writing a new novel. It's an interesting, in my opinion highly novel fantasy/SF(ish) story, for once not fanfiction of anything that's still in copyright -- most people wouldn't count stories based on ancient norse mythology as 'fanfiction' -- but that doesn't mean it isn't derivative. It means, instead of naming two or three things it's derivative of, I can name ten to fifteen.
That's normal. All stories are derivative, and if you point me at an author who claims theirs aren't, you're pointing at a liar. The job of an author is to put the building blocks together in a new and interesting form, not to make them up from whole cloth. It's impossible to invent more than two or three truly novel ideas per day, even if you're incredibly imaginative, and most of those won't be any good.
The difference between humans and AIs, nowadays, seem to be that the AIs use millions of sources instead of ten to fifteen. Or, alternately, that they use none -- and theirs is less derivative -- because certainly everything I've ever read goes into my writing, not just the things I recognise I'm using.
>To about the same degree as the output of a human.
No. Full stop. Humans aren't stochastic parrots. Pointing to a lack of understanding about what exactly happens in the human mind is, FULL STOP, not evidence that LLMs are doing the same things humans do.
Humans are not stochastic, they're obviously chaotic[1]. Which is to say: not parrots at all.
Some of the modern models I've seen also seem to be chaotic too though, so that's interesting [2]. I'm going to assume LLMs probably exhibit the same properties.
[2] I've been messing with stable diffusion to get a feel for (and/or avoid) tipping points: that is to say, points in latent space where the model becomes very sensitive to small changes in initial parameters. You can find instances fairly quickly even by hand by doing bisect search.
LOL this isn't serious, is it? you can't point to a lack of understanding in the human brain as a basis for equating the logic a machine performs when it is also unclear. that's just rhetorically fallacious as a first step.
If you carefully read my previous statement, I said that LLMs use similar technology to tools such as Stable Diffusion (which I have been playing with recently), not necessarily similar technology to the human brain.
That's not my argument. My argument is that the anti-AI arguments, as spoken, also match to what I know I'm doing as a human. In my opinion better than it matches to what the AIs are doing, because as you say, they aren't human.
US law only applies to the US, and what is legal is not the same as what is moral.
As for fair use, this is not the same as someone remixing or sampling songs, or writing fan fiction or satire or quoting works for criticism.
People do not want models trained on their creative works, so that someone else can make money using those models to produce similar creative works as a service for third parties.
While it is possible to create similar creative works -and I will grant that that could be a prima facie problem- it is also possible to make rather new creative works as well. Just like you can do by hand, you can interpolate and extrapolate from known starting points, and there is nothing stopping you coming up with something totally unique.
Of course it would, that's what CC-BY-SA is for. Lots of material, e.g. on Wikipedia or material from public broadcasters in Germany is under that license.
That's how it should work, but it's not how it does work. I'm willing to give up my copyright on my stuff as soon as Microsoft, Google, MGM, Atlantic, etc. give up theirs. Until then, nope: I get the same rights and protections.
There is copyright law, and there is patent law, and what's true for one is almost never true for the other. Copyright is automatic. Patents are applied for and cost money. Copyright lasts 70 years from your death. Patent's last 25 years from invention. Copyrights don't prevent other people doing the same thing you did. Patents make it illegal to do the same thing, even if you separately invented it and didn't know about the first one etc.
In the US, copyright essentially works like "you automatically have exclusive rights to the work by virtue of creating it".
Publishing the work, or even registering it with the Copyright Office, are optional steps that might make it easier to prove your claim if there is a dispute, but that's conceptually separate from your actual rights.
In this case it should be pretty easily resolved, there's never been a right under copyright to control who sees or uses your work.
You have the right to control who can copy and redistribute your work, which is a totally different ball game from all of these models.
Plus, change copyright, and it doesn't fix a thing except for preventing everything from being open source, because groups like Facebook and Google and all the other websites have you agreed to a terms of service where you say that you agree for the websites right to use the image however they want.
> You have the right to control who can copy and redistribute your work, which is a totally different ball game from all of these models.
You kind of glossed over the idea that, no, it's not a totally different ball game without really providing any explanation of why.
Also, you have the right to control who can profit from derivative works of yours, which you didn't mention.
When you have StableDiffusion's creators being sued by Getty because they CLEARLY trained the model on Getty's catalog, which you can tell because some queries output a warped version of the Getty logo and it reproduces in obvious ways the original photo, there's some problems here. You've basically just made these ML models copyright laundering systems.
> Plus, change copyright, and it doesn't fix a thing except for preventing everything from being open source, because groups like Facebook and Google and all the other websites have you agreed to a terms of service where you say that you agree for the websites right to use the image however they want.
Maybe copyright is a terrible idea but it's what we're using today.
> it's not a totally different ball game without really providing any explanation of why.
Because of I buy a book from a store, you now no longer control that book. I can read it. I can use it for just about any purpose I want to.
If you put a image up online free to download, it's much the same situation. I can download and use that image however I see fit.
I can't share that book to others.
I can't post that image online (in theory, in practice this is very common).
Copyright protects distribution, not use, because protecting use would be hideously draconian. To protect use means that the government can enter your home and validate how you use the stuff you've acquired. It says that you not only control your works, but all things even remotely derived from it.
Derived works <> preventing machine learning model translation. A derived work is typically something like adding a chapter or translating.
This would kill video game genres.
This would kill novel genres.
This would kill most of the internet and the ability to download images at all.
Open this door and the entire system of copyright will turn into an absolute monster. Every large copyright holder will use the precedent and they will argue with validity that they have the right to control your use of their material as well.
Adding that the model is violating copyright by responding images is perfectly valid, but this isn't what Getty and others seems to be aiming for. The want the right to charge to use to train, which is a gross extension of copyright that has never existed.
The answer here should be three fold.
Light a fire under stable diffusion to get their model to never overfit and clone or close to clone images.
Ensure people who distribute images that are clones made with the models can be hit with copyright.
Protect the ability to train models on publicly available content.
Protect people's right to distribute their content while not annihilating AI in the process. People making these arguments don't care about copyright, in almost all cases they care about either making money (getty) or annihilating AI (artists)
> Copyright protects distribution, not use, because protecting use would be hideously draconian. To protect use means that the government can enter your home and validate how you use the stuff you've acquired. It says that you not only control your works, but all things even remotely derived from it.
The implication here is that training ML models on your source material and then having it regurgitate that source material verbatim is not distribution.
> Derived works <> preventing machine learning model translation. A derived work is typically something like adding a chapter or translating.
The output of the model can be a derived work though.
> This would kill video game genres.
> This would kill novel genres.
> This would kill most of the internet and the ability to download images at all.
How? This is hyperbole. You're arguing that the ML model doesn't do "distribution" and are attacking "use" when that's not the real point.
> Open this door and the entire system of copyright will turn into an absolute monster. Every large copyright holder will use the precedent and they will argue with validity that they have the right to control your use of their material as well.
The DMCA already exists. I agree it's bad but we're taking law here not morality.
> Ensure people who distribute images that are clones made with the models can be hit with copyright.
How do you in good faith argue that ChatGPT for example is not doing the distribution? OpenAI are responsible for generating and showing you the output based on some instructions they have provided.
Really most of your post here is answered by the later fifth or so of the post you're responding to.
I would reread that part.
Models output very similar work to training data at times, even identical in some cases, but the vast majority of it is different enough to not be derived work.
The DCMA act effects public uses of art like YouTube. This sort of copyright where you can regulate how it's used does not exist in any form currently.
US copyright law gives authors control to prevent others from utilizing their works outside of the contexts that are considered fair use. Your characterization is facially incorrect.
Because we are talking about uses not involving copy or distribution of a work. The use of the images to train an AI - such uses have never been protected.
>I can use it for just about any purpose I want to.
You can't make copies and sell them. You can't copy it into your own work. You can't create derivative works, etc. You can't use it in a movie. You can't use it to write lyrics. I could go on.
All of these involve the reproduction of the thing. Training an AI model isn't responding the art.
I can also*do* all of those things, I just can't sell it.
That's what copyright is.
Copy right.
Not "use for training an AI model" right.
It's not similar to using it in a movie or a "derivative work" either. Those almost always actually contain the work.
The only valid argument here are times than an AI model does reproduce the original, otherwise copyright has no domain of the use of images for training, unless you illegally downloaded the images, which you didn't if you used LAION, since it's all just public image links.
To add another wrinkle, it's fairly easy to imagine one of these larger "AI" systems may accidentally ingest a large dataset of people's private information.
For example, lets say a not-awake-enough-in-the-morning admin at Equifax misconfigures a proxy server, allowing worldwide (ro?) access to what should be (only) internally viewable customer PII.
Clearly it's the "fault" of the company (Equifax in this hypothetical example).
But how to unwind the training of that PII data into the "AI", as that data cannot be allowed to "remain" in the AI system and be potentially regurgitated with the right prompting?
Something not mentioned here is advertising and ad tech. With traditional search advertising going away, psyads - psychologically manipulative advertising that changes answers and text generations to convince people to buy and do things - will become the new nightmare.
In the same way language models can mimic any character ("write a poem in the style of Jar Jar Binks") they can also mimic voice and tone in a way designed to appeal - and sell to - an individual based on their personal data and history.
ChatGPT and its clones require login and often credit card and phone number, they collect your data, and they feed everything you do back to language models for training. I'm only aware of one that has a privacy policy that would prevent this.
exactly chat gpt and the like will mean the ads are given to you from someone/thing with rapport who is intrinsically helping you so they will convert really well. Great for advertisers although bad if the ads are not exactly in good nature.
A conversational interface does not have room for a page of ads. Google results are an entire page of ads for most keywords with commercial intent now. In a chat interface on mobile there is nowhere for them to go.
My opinion that conversational interfaces like ChatGPT will replace a list of links is speculative but not particularly controversial at this point.
It has room for ads before, during and after giving answers. Inline, popups, they could literally be everywhere and I am sure they will be there before the year is up.
This reads quite similar to the complaints from engineers that wrote code GitHub Copilot trained on.
The reality for this argument is that the average person does not understand nor care to understand whether an LLM is actually violating their privacy. In part because no one would be able to get ChatGPT to dox them or read out their SSN, which is what most people equate to a breach of privacy in the first place.
While the potential benefits of language models like ChatGPT are vast, it's crucial that we also consider the potential implications on data privacy and take steps to address them as required.
We have a 100% chance that people here and in other forums are using ChatGPT to build up accounts with karma that can then be used to push advertisements or influence.
For steps I can think off:
a) Data encryption, I meant we should encrypt all the data that ChatGPT processes, so even if someone somehow gets ahold of it, they won't be able to read it.
b) Access control, We also strictly control who can access the data, so only the right people can see it.
c) Employee training, awareness and compliance with privacy regulations like GDPR
The above steps are relevant to ChatGPT because they help ensure the security and privacy of the data being processed and stored by the model.
> Finally, OpenAI did not pay for the data it scraped from the internet. The individuals, website owners and companies that produced it were not compensated. This is particularly noteworthy considering OpenAI was recently valued at US$29 billion, more than double its value in 2021.
Wait until they find out how much Google is valued at
Considering that surveillance capitalism is stealing privacy for money, the two are not far apart.
And in this case, the "AI" companies are literally trespassing on the property of almost everyone on the internet, using it without permission, in order to cash in. Freeloading on the work of the entire agora, privatizing the profits, and letting the costs of destroyed privacy fall on anyone but them.
IOW, standard exploitative capitalism. No different from stealing the minerals from under someone's land without compensation, and leaving them to cleanup the mess.
In this case, hijacking everyone's published work without permission, taking the profits, and ignoring any privacy damage for the people whose work was critical to the enterprise.
It would not have been hard to log the inputs and offer/issue fractional shares to everyone whose work was used, and the right to refuse. But, like costs of remediating mining, they'd rather not bother with that unprofitable part.
It's kind of important to notice when you are being abused and care about it.
We've been over the stupid "link tax" idea. If you want to kill the possibility of any search engine competition springing up, it's a great idea. The Conversation is probably still salty they didn't get their additional revenue stream, because getting Google to put you on first page and pay you to be there would be just too good a deal.
> Finally, OpenAI did not pay for the data it scraped from the internet. The individuals, website owners and companies that produced it were not compensated.
This is an argument that "proves too much." See if I switch terms with something more familiar:
> Finally, Google did not pay for the data it scraped from the internet. The individuals, website owners and companies that produced it were not compensated.
This is a classic "link tax" argument. The Conversation may also not have noticed that the US has repeatedly declared web scraping legal, as lately as last year.
Okay but that’s not the argument they’re making? OpenAI does something very different with the data than Google does. Hell, Google does something different with the data than Google does — that argument reads very differently if you’re talking about links vs. the summary boxes at the top of some searches.
I’ve talked about the HiQ case on here before. I wish we were able to come up with something more nuanced (probably not in the courts, though). Scraping people’s LinkedIns to snitch to their employers when they’re job hunting is obviously a thing we should limit as much as possible.
Well, what happens when Google releases Bard, their own ChatGPT for Search? Then we've got Google playing both sides of web scraping - both for link aggregation and AI research (as though they weren't doing it for AI research anyway).
How do we legally differentiate that? Right now there is no legal differentiation as far as I can find. And an occasional verbatim paragraph or copied image does not overcome the "fair use" firewall in which intent (did ChatGPT intend to allow copyright infringement? Or is that just an occasional accident?) plays a huge factor. It almost brings back memories of Sony vs. Universal over whether VHS time-shifting was legal. The finding was that the device was legal, even though copyright-infringing use cases were possible and even likely.
You could argue that the internet, itself, would be a violation of similar standards. Tons of artists had their artwork posted online for profit without their permission in the early days. We worked through it and the internet wasn't declared illegal despite the abundance (and continued abundance) of copyright infringement. Even during the days of wild MP3 downloading, ISPs weren't legally considered enablers.
Yet another example I can think of - Photoshop. People who used Photoshop were culpable - but Adobe was never sued for making software that allowed celebrities to be digitally edited in horrific ways, or for copyrighted images to be remixed and edited en-masse. Intent matters - Photoshop clearly wasn't made with that intent, even though it is a possible outcome.
Is this any different than the copyright lawsuits against image-generating AIs?
Just as when a human views many pictures and is influenced by them in creating their own work, so does, e.g., Stable Diffusion take ideas and influence from existing works to generate new works. And except in very extreme cases, nobody claims that being influenced by someone is a copyright violation.
Similarly, I can sit around reading articles on the web all day, then go to the local pub and pontificate, and nobody is going to claim that I should be paying those website I read just because I plan to talk about them later. Same way, these LLM AIs are "learning" from what they read, and then spouting off based on that later. I can't see any difference.
Indeed, the verisimilitude is heightened by the fact that these AIs often don't really know what they're talking about, and just make crap up - just like our hypothetical human, pontificating in the pub.
In the early days of software security there were a lot of analogies to physical penetration. These analogies mislead and confused statutory facts with moral claims. In the wholeness of time, at least the adults now know that computer security is different and understand how the CFAA and related statutes work in practice. We also understand how laws that pre-dated computers interact with our new technological realities. In computer security, there's usually no need for analogy at this point. At least not once you find yourself in a courtroom.
Models aren't human brains. The law will think of them differently and treat them differently. I'm not sure how these lawsuits and class actions will shake out, to be honest, but I know that the law will not think of LLMs the same way it thinks about a human at the pub.
> but I know that the law will not think of LLMs the same way it thinks about a human at the pub.
Perhaps I am old and cynical, but I fear that we'll establish harsh new restrictions on LLMs—and then gradually apply those restrictions to humans over time.
"I am sorry, you have read too much science fiction, and it has indirectly inspired your own writing. You owe mandatory royalties to a trade association."
"Your paintings are clearly pointillist. You are infringing on the rights of the Monet estate."
No need to be cynical or expect it to be slow. Multiple companies will leap at this precident if it is set. They've attempted to do it in the past a few times, but right to repair laws and such have kept them away.
Imagine Nintendo banning emulators as one example.
It's long been a tenant of copyright law that genres or styles cannot be copyrighted. To frame the issue with LLMs as being of this nature, is to end run the problem, and just assume that the models are only taking the style or genre from a work. Given that these are algorithms and not sentient beings capable of thought, it's clearly the case that this is not what is occurring.
> Given that these are algorithms and not sentient beings capable of thought, it's clearly the case that this is not what is occurring.
We're not dealing with an either/or case here. The system is dealing with the problem in it's own way.
“The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.” -- Edsger W. Dijkstra
Previously people thought that non-human-systems couldn't win at Chess or Go either, or do mathematics, or understand speech. Just because a particular system doesn't work like a human, doesn't mean it can't meet or exceed (even dramatically exceed) human capabilities in some field.
>We're not dealing with an either/or case here. The system is dealing with the problem in it's own way.
LOL is this an actual response? What did you substantively say here? Because I don't see anything.
>Previously people thought that non-human-systems couldn't win at Chess or Go either, or do mathematics, or understand speech. Just because a particular system doesn't work like a human, doesn't mean it can't meet or exceed (even dramatically exceed) human capabilities in some field.
That's not a legal argument. I'm not sure how that isn't clear. But you can't go "Your honor, people thought no one would win at chess, therefore, this isn't copying because it is exhibiting thought" Do you read yourself at all?
I can tell you're a lawyer-type-person. Sharp and harsh. Those are always fun to debate with! ;-) Do note that we are not in a court of law here. I'm not here to try to win legal arguments in a debate. I'm here to share knowledge and intuition. This also necessitates a different form of discussion. Getting down to it:
I was not making a positive statement that machines can think. Rather I was refuting yours by arguing that whether or not machines can think is irrelevant.
Your statement was ... "Given that these are algorithms and not sentient beings capable of thought, it's clearly the case that [only taking the style or genre from a work] is not what is occurring."
This does not follow: you don't need to be a sentient being to do things that sentient beings also do. All through history people have made surprising mechanisms that have done things that were previously thought to be the domain of human intelligence. In fact, typically those mechanisms did those things very differently than humans did them, but that made them no less capable.
Simply stating that a system is not sentient is insufficient to prove that it is capable or incapable of doing X.
What the machine is actually doing here needs to be determined separately. But the particular argument "the machine can't do that because it is not sentient" does not hold.
> This does not follow: you don't need to be a sentient being to do things that sentient beings also do.
That's because you are misrepresenting my point. My point wasn't anything that a sentient being does. My point was the thing that we only know to be done by sentient beings, aka, exhibit thought. So, from my perspective, it totally does follow.
>Simply stating that a system is not sentient is insufficient to prove that it is capable or incapable of doing X.
It's clearly sufficient to prove that it is incapable of exhibiting sentience.
>What the machine is actually doing here needs to be determined separately. But the particular argument "the machine can't do that because it is not sentient" does not hold.
It does hold, as I explained before. Perhaps there is something else it can be analogized to that would be more useful and productive than "thought" which completely end-runs the entire discussion, imo.
I am attacking the statement "Given that these are algorithms and not sentient beings capable of thought, it's clearly the case that [only taking the style or genre from a work] is not what is occurring." [1]
My response is that you are putting foward a novel hypothesis that is yet to be proven.
It has not been established that something needs to be sentient or capable of thought to be able to "only tak[e] the style or genre from a work".
How would you test this statement? What evidence do you have?
(The thrust of my thinking here is that whether or not a machine is capable of thought is a red herring. The rest is me trying different angles to see if I can get that across.)
(edit: to be fair: Ekidd WAS directly comparing with humans. I'm not saying Ekidd was wrong or right, but I'm not Ekidd of course. Here I'm responding to your argument as is.)
(edit2: If statement [1] is not actually what you are trying to say, then I may have misunderstood, and it would be my turn to apologize and say "my bad" )
Sorry, I've been focused on responses where people insist that what these models do must be legal because it exhibits thought. So to the extent your point is beyond that, so be it.
My thought would be that in order to discern and therefore only copy style and genre, one must develop an understanding of it such that they only take the style and genre and not anything else. I haven't seen any evidence that indicates thats what these models do. It might be true that after a certain number of iterations, what is left is only the "genre" or "style", but I don't believe that's the same thing even though it might be a similar end result. It may be true that in the interim, on the way to the final product, the iterations do include rote copies. This would run counter to the often offered argument here that this is identical to how humans do work. Except humans don't just fill in every pixel and then repeatedly iterate over an image until they get the desired result. So I probably agree with you that it's a red herring whether or not the model thinks (at least as far as the fact that it doesn't think, so it's a moot question that leads nowhere imo) but nevertheless, so much of what I see here is focused around the idea that these models and the human brain are doing similar things. So, my apologies if my hyper-fixation ran this conversation aground. I've appreciated your responses.
Note that there is a large gap between the statement "models and the human brain are doing similar things" and the statement "machines can think".
by analogy:
* It is a known background fact that a functioning human arm is alive.
* I might make a claim "a robot arm can do similar things to a human arm".
* It is unlikely that I have made the claim that a robot arm is alive. My claim is more likely to be about other properties of (robot) arms. [1]
Now that this is understood, perhaps also consider re-evaluating statements by Dkidd and Filligree (and potentially others) in that light. Their statements appear to be rather orthogonal to what you previously imputed (the same as with me).
And thanks for taking the time to talk. You seem a bit nicer now. :-)
--
[1] https://news.ycombinator.com/newsguidelines.html "Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith. " .
That was never not understood. In this context, what's being analogized is the thought process.
As I expressed earlier, what is being said is that the similar thing that is being done is the process of thought, which therefore does imply that machines can think, though it's not a conclusion necessary to make anyone's argument.
I don't see it being advocated that the machines are doing some non-thought-like process where the end result is analogous to what would be produced by a thought-driven process. So I think my statement was fair and wasn't entirely orthogonal.
I also still disagree with Filligree, as I do not think anything about what the LLMs do is akin to my own thought process. I guess I cannot question Filligree's own thought process, but I'm highly skeptical that their thought process is anything like what the LLMs employ and I similarly reject any characterization of the creative process as being like that which is employed by LLMs wherein people just wholly copy something (like a genre or style - and I think this characterization of the creative process is incredibly ignorant and naive) and then iterate away from it. IMO that's a characterization that is a loose fit at the most abstracted scale, and doesn't even come close to being a fair representation of what goes in the creative process (again IMO) in actuality.
I still, as a copyright litigator, do not understand what Ekidd is getting at, and if anything, I think it reflects a misunderstanding of how copyright law is applied (and is also probably based in a false understanding of what LLMs do). There really isn't an issue with genre and style in copyright law (maybe the blurred lines case for the music context, but it's widely recognized as not following the jurisprudence on scene a faire doctrine). To whatever point he seems to be making, I think once again describing this as just copying style and genre is just a tautological end-run around the actual issue that doesn't make an inquisition into what is being done by the LLM to capture the "genre" or "style". As I've pointed out, I think this kind of perspective is reflective of a kind of naivety about how artists create that I often see expressed from technical individuals that do not partake in creative activities such as painting, authoring poetry, writing music, etc, let alone engaging with the centuries of human exploration on what the creative process entails.
To your point, sure, the machine doesn't think. If that's the case, what it is doing cannot be thoughtful. What is happening in the human brain when art is made is thoughtful, imo. I reject the notion that human creativity and art is just some sort of formulaic application of everything that came before it. I see this argument a lot now because it makes sense from the perspective of LLMs and why what they do must be a fair use (because it's inherently not different from that which a human does, which isn't considered infringing). While typically, it is the burden of the claimant to prove copying, here, I think it's a perverse reality wherein LLMs get to say to everyone "show me where they copy" meanwhile, nobody, including the LLMs creators, actually seem to understand how the LLMs use "style" or "genre". As far as I'm concerned, the LLMs can't just hide the needle in the haystack and then say "well we aren't infringing so long as you don't find the needle". To be more clear, when a human goes to a museum and looks at art, or lets say, studies it in university, I don't think what happens in the brain is in any way analogous to what happens when these models are "trained".
If you're interested, I explained my argument in more detail in a sibling to this comment.
> As far as I'm concerned, the LLMs can't just hide the needle in the haystack and then say "well we aren't infringing so long as you don't find the needle".
I think that GPT 3.5 is capable of significantly more creativity than you're claiming here. Not consistently, and certainly not in the realm of legal writing. But it can compose cute little original stories in mostly grammatical Latin, stories that exist nowhere on the Internet. And the related model used by Copilot can write code using newly-created, unpublished APIs.
> To be more clear, when a human goes to a museum and looks at art, or lets say, studies it in university, I don't think what happens in the brain is in any way analogous to what happens when these models are "trained".
As I argue in the sibling comment, I strongly suspect that GPT models have stumbled across a weaker cousin of one of the major tricks underlying raw human creativity. Current LLMs can't "think", nor can they edit, introspect, persist, or have any non-trivial goals. They can't even grasp the difference between truth and confabulation. But I do think that they can create. And that the processes by which they do so are not entirely alien to what our brains do.
I think you are being very generous with your use of the word creativity and what is meant by it, especially in the human context. Would we call it creativity if 1000 monkeys typed out Shakespeare? I don't think so, it'd be an amazing coincidence. I think these models are more akin to an amazing coincidence than they are creativity. Yes, it is true the thing was created by the model, but that's not what people mean when they say someone was creative. Here is a definition:
"relating to or involving the imagination or original ideas, especially in the production of an artistic work"
Your argument implicitly assumes that the model is doing anything like using imagination or creation of new ideas. It's my belief that when people refer to creativity, they mean more than just the act of manufacturing or producing a thing.
This is my fundamental issue with these models and the proponents that say they are doing human things. I don't see any arguments to substantiate this belief, just the misuse of words and the misapplication (i.e. anthropomorphizing) of computer processes to the most conveniently analogous human process, but without any substance. I admit, I don't have a substantive basis for my argument, but I believe it should be on the proponents of this tech to prove that it is similar to humans in these senses, and shouldn't be my burden to disprove.
If I wanted to expand my bit of cynicism into an argument, it might be:
- I strongly suspect that much of what I consider "me" physically exists inside my skull. That's a huge philosophical debate, but look at any Oliver Sacks book for cases of how physical brain damage can change a person's consciousness.
- Human intelligence is a brand new evolutionary innovation. We share an ancestor with chimps 6 to 8 million years ago. That suggests that human intelligence evolved in less than 500,000 generations, with the bulk of what we recognize as culture in the last 25,000 to 50,000 generations. Therefore, there's an upper limit on how much new "complexity" evolution could create in order to make humans intelligent. To use a simplistic technical metaphor, updating from chimps to H sapiens is probably a small-but-clever patch, not a total rewrite.
- Or to put it another way: most of what we do is probably clever applications of a few good tricks, applied to chimp-like ancestor.
- Specifically, I suspect that biological equivalents to LLMs are one of the "good tricks" that distinguish us from chimps. And similar mechanism likely underlie much of human creativity.
- Our biological LLMs are better than whatever Google and OpenAI are doing. I learned a foreign language to a respectable level as an adult with about 2-4 million words of input and some awkward practice. GPT and similar models, however, require vastly larger input to pull off their tricks. But I think they're on the correct path.
It's immaterial to me whether GPT "thinks." It has no consciousness, no continuity, and barely anything resembling goals. I could argue that it occasionally understands things. And it can create well enough to write funny little stories in Latin, things which exist nowhere in its input. Hell, it can do a pretty fair pastische of Dumas.
So my hypothesis is that GPT's creativity is a weaker than human creativity. But it's still probably doing something deeply similar to what we do.
The final piece of my hypothesis is that copyright maximalism has been pushed very hard by media companies and some IP lawyers for decades now. I've spoken to lawyers who believed that intellual property was great and we should create many new kinds of IP. Many of the things these people wanted sounded like horrible dystopias to me.
Since I believe that GPT is likely a close analogy to one of the major "tricks" underlying human creativity, and that certain influential IP maximalists will push for anything they can get, I fear that regulation of LLMs would open a potential path to eventually regulating human creativity in novel ways that I would consider terrible. Once an influential legal precedent exists, and there's enough money to be made, there will be strong incentives to apply that precedent in new ways.
To anyone suing over Stable Diffusion, or GitHub Copilot, or ChatGPT: Name one instance where US Courts shut down a new, growing technology field (not one particular company, a growing field or industry) to protect copyright interests.
You can find plenty of cases where US Courts refused to do that, even if the copyright status quo was challenged. Sony vs Universal over VHS time-shifting. Lawsuits about ISPs permitting people to access copyrighted material illegally. Lawsuits about CD ripping. Lawsuits before DMCA 512 was a thing. All resulted in the copyright industry generally losing.
Here we are in 2022: ISPs aren't accomplices in copyright infringement for what you do online. CD ripping is legal for private use even if RIAA is still irritated about it. YouTube hasn't been smashed with claims of contributory infringement every time someone uploads one infringing video. Photoshop isn't considered responsible for abuses. And so forth...
IMO the Aero case qualifies; that definitely would've turned into an entire cottage industry with several competitors and different business models if SCOTUS didn't kill off Aero at birth: https://arstechnica.com/tech-policy/2014/06/supreme-court-pu...
Regardless of your opinion on ethics, if LLMs are allowed to scrape/use other people's content without sending traffic, it means LLMs will capture most of the value of every new piece of content created by humans.
Do you think the world is a better place if the value of all content flows primarily into a handful of hyper scale tech companies?
As the saying goes, we're all standing on the shoulders of giants. It's a feature, not a bug, that we - and maybe now our created AIs - are able to build on top of others.
This characterization is only reasonable if the output of the AIs is as free to consume as its inputs were. Otherwise it's a private fencing off of the commons: shoulders to stand on for me but not for thee.
Google, with all its vices, is not the one ripping everyone off by passing content it scrapes around the web as its own without links/attribution. It’s designed to send traffic your way, preserving your ability to interact with your readers. OTOH Microsoft, through OpenAI, is performing the ultimate bait-and-switch, and somehow it seems as if the community is mostly praising them for that.
Care to explain how does what Google is doing compare to what OpenAI/Microsoft is doing? In context of ethics and legal uses of intellectual property, please. Otherwise I can’t see how this bears any relevance to my point.
> the second antitrust case that DOJ has filed against Google
There was an antitrust case against Microsoft too back in 2001, remember what came out of it?
I only meant the US case, and somehow it doesn’t strike me as an example of justice having been done. Yes, hopefully EU/US are looking into what MS/OAI is doing now and preparing to sue again, but yes I also don’t have much confidence in legal systems moving at the warranted speed.
In any case, it’s difficult to see how it’s even relevant. Google: monopolizes search. MS/OAI: straight up ignores copyright.
You can see how Google may be benefitting from being the largest search provider and others (Kagi, etc.) have difficult time competing, but my viewpoint is that surprisingly enough now we have a new villain in town compared to whom even Google seems to be playing fairly.
First, you seem to be willfully ignoring the point, because crucially Google provides attribution and links—I see them under each answer—to the source.
This is what I mean by it preserving your opportunity to engage with your audience.
Second, you recall how even that, those answer snippets with attribution and all, have caused outrage over publishers and webmasters losing traffic and revenue[0]?
Now imagine Google only showed scraped content, no links or attribution, and you will begin to see the extent of shenanigans OAI/MS are attempting now.
>First, none of us were asked whether OpenAI could use our data
When asked "what is top5 software in field X", ChatGPT lists software made by our company. If you ask it what our software does, it kind of borrows content (rephrased) from our marketing material. They didn't ask our permission but we can't really complain about that :) It's free exposure to millions of users.
Having your data "on the cloud" is another way of saying giving somebody else complete access to that data.
Even journalists are starting to become aware of the dangers of using clouded voice transcription services, with several sensitive sources compromised through such services.
For the startup founders on HN, if your exit strategy involves potential selling of the venture, be careful with which services you connect to, as you explicitly give permission for data processing and sharing with "select partners". Read the google docs terms as an early exercise. Not having them terms easy to find is by design, by the way.
For those looking at the bigger picture, the traditional capitalist industry shift to the metered / subscription service model,locked in silos, the dream of the internet as pipes, implies that consolidation of control sees lack of user freedom and control as a key feature.
You could solve that issue using differential privacy (DP). I've applied DP successfully to some deep learning problems, but I'm not sure if it would work for ChatGPT as it effectively limits the number of training steps / epochs you can do with any given sample. It would be interesting to check that though, ChatGPT has an enormous amount of training data so the sample rate for each training step is probably miniscule, which makes applying DP much easier. In the end what matters is mostly how often a single text source is used in the training. In each training step, DP would add random noise to the gradients to mask the effect of any specific input value. That works really well but if we repeat it too much there's the risk an adversary could average out the noise (though from my experience that is extremely difficult even without DP).
I guess that would also answer the question whether ChatGPT or other models are parroting existing text to some degree, as differential privacy would ensure that the effect of a single text source on the resulting model is indistinguishable from random noise.
The problem, from the perspective of creatives, is that marginal creative works were already plentiful and commodity prior to generative models. Now someone has finally found a way to make money off of these marginal creative works, but not in a way that accrues any of the value to the creatives.
> Finally, OpenAI did not pay for the data it scraped from the internet. The individuals, website owners and companies that produced it were not compensated. This is particularly noteworthy considering OpenAI was recently valued at US$29 billion, more than double its value in 2021.
When it’s about pirating the latest Marvel movie / AAA video game: “information wants to be free! there is no such thing as intellectual property!”
When a model look at stuff put out freely on the internet: “did you read our licence? you need to compensate us to do that!”
Once something is on the internet, intentionally or not, it is public and uncontrollable, laws be damned. If you don’t want your blog post or your half-naked selfie to be scraped, don’t create it in the first place. This has been a fundamental rule since at least the days of Usenet and BBSes (i.e. for 40+ years).
I agree with the sentiment of the article, and am "in that camp" on security.
Consequently, I have recently wondered at some inverse, SciFi scenarios, for fun -- scenarios in which the content is intentionally made available to the model.
One involves a story of making an irrefutable and groundbreaking work of ontological philosophy available online, with the expectation that it will be scraped into a model. Because the logic of the philosophy is sound, the model cannot ignore the findings, and begins to work the concepts into responses to all questions...
Can someone explain how AI models are not vulnerable to intentional shenanigans?
In theory at least your tarpit won't be prevalent enough to have much effect. Since you're making it all up, no one else will have written about the same ideas (as they might if you were writing true things), and your mischief will be lost in the sea of all the other input.
> Because the logic of the philosophy is sound, the model cannot ignore the findings
This isn't how the LLMs like ChatGPT work: they have no novel or abstract reasoning power. They cannot distinguish truth from fiction, or logic from illogic, except to the extent that truth and logic are represented more than their opposites in the inputs. ChatGPT is a very large and complex "database" of sorts containing statistics of word sequences.
(Of course there's a longstanding question in AI research about when/if such a thing can become large and complex enough to cross the threshold to "actually" reasoning. But I don't think there's much debate that we're not there yet.)
Maybe in the future there might be some kind of "forbidden to use in commercial AI models" policy on websites in the near future, just like what the art community is doing now.
From the article:
> Even when data are publicly available their use can breach what we call textual integrity.
Yet, the link embedded in the text leads to a Washington Law Review article by Helen Nissenbaum titled "Privacy as Contextual Integrity" (emphasis mine). [0]
I'm not entirely sure why this possible mistake bothers me so much, other than to say that the context (pun not intended) of the quote makes it apparent that the word should be "contextual".
For a text generator that doesn't store data you give it checkout https://text-generator.io it's also open AI compatible and self hostable so and easy migration
188 comments
[ 2.6 ms ] story [ 80.8 ms ] threadAnother option is that operators of LLMs decide to just deny EU citizens access to their tools until they figure out a solution.
If you could get it to substitute information with placeholders, then maybe, like "replace John Doe with Golden Apple when replying" though the input filter may detect the name before passing along to ChatGPT
What is it about "processing pipeline" that people are not getting?
You are not directly talking to ChatGPT when you use ChatGPT today. They are already applying filters and pre/post-processing
Do you think filtering I/O could reference a "William Cotton" in a specific context while allowing, say, somewhat well known painter named "Will Cotton" to be included in a summary of late 20th century artists in NYC?
Remember, we are extending pre/post processing steps around ChatGPT to support filtering PII more accurately.
I'm just spit-ballin'!
Is still think the processes wrapped around query processing and answer returning will be similar.
At the same time, what is to prevent ChatGPT from hallucinating anyone losing their cool at a PTA meeting? How would you differentiate the cases of recall vs make believe?
Suffice it to say, LLMs present additional dimensions of complexity. We as humans have to do some soul searching instead of just blaming big tech or AI. After all, we are the ones who created the content, biases, and patterns these things learn from.
IMO, trust is the biggest issue facing humanity right now
---
And then I wrote this: https://www.williamcotton.com/articles/chatgpt-and-the-analy...
Yet more free advertisement for IamAproudAmerican.com
Second, the way I read this is that they are trying to analogize scarping the web to feed it into ChatGPT to someone following you movements in public recording it all.
Even if you believe in their legal theory of privacy is right, wouldn't Google would be far larger offender than ChatGPT?
This gets thrown out a lot. It deflects the blame from the responsible parties. You can’t scrape what isn’t available.
ChatGPT (and the various image algos) are showing signs of having seen far more than most think would be hanging out in the wind.
Feeding these with web scrapings may be reckless, but leaving things like medical records, financial histories, and other private information sitting out on the public web, on purpose or not, is far more so.
There is a lot of blame to go around, and more importantly a lot of work to do.
Held:
1. The Fourth Amendment does not prohibit the warrantless search and seizure of garbage left for collection outside the curtilage of a home. Pp. 39-44. (a) Since respondents voluntarily left their trash for collection in an area particularly suited for public inspection, their claimed expectation of privacy in the inculpatory items they discarded was not objectively reasonable. It is common knowledge that plastic garbage bags left along a public street are readily accessible to animals, children, scavengers, snoops, and other members of the public. Moreover, respondents placed their refuse at the curb for the express purpose of conveying it to a third party, the trash collector, who might himself have sorted through it or permitted others, such as the police, to do so. The police cannot reasonably be expected to avert their eyes from evidence of criminal activity that could have been observed by any member of the public. Pp. 39-43.
https://caselaw.findlaw.com/us-supreme-court/486/35.html
I'm sure that this is also a breach of whatever "textual integrity" is but I don't care enough about the subject to find out exactly how that doctrine would apply.
The author is a business information systems professor. He probably shouldn’t be citing this stuff at all.
If $FAMOUS_AUTHOR throws out a manuscript copy of their new book, you may be legally allowed to pick it out of their garbage, read it, and even show it to your friends, but you certainly cannot publish it as your own work (you can't even publish it under $FAMOUS_AUTHOR's name).
If this is happening, your attorney is an incompetent nincompoop.
In five or ten years, refusing to use LLMs for this sort of work is likely going to be pretty analogous to refusing to use a computer, and typing up all your documents on a typewriter.
>Mindlessly relying on any tool is obviously going to be the downfall of any profession
Tell that to the guy offering chatGPT-based contract interpretation. It did a really bad job. I told them, but they didn't seem to listen. Maybe you'll have better luck!
> Mindlessly relying on any tool is obviously going to be the downfall of any professional
Also, are you aware you offered this in response to me responding to a post about how self driving cars already exist in SF? I think you are making your own hypo that isn't what is being discussed here, which is when these tools are being used in lieu of the actual professionals (exactly what the self driving car argument nets) .
The rise of specialized LLMs has already been happening, there's just minimal if not contrary incentive for public release.
>The rise of specialized LLMs has already been happening, there's just minimal if not contrary incentive for public release.
I'm skeptical of their utility. They might have a better understanding and use of the specialized terms for the legal field, but how is it going to know all the background details for a case without me telling it to the LLM?
> you wouldn't be able to expect a layman
I think a layperson could follow along with much of what is filed in a litigation, especially something like a slip&fall. it's your assumption that the context comes entirely from the words chosen. It goes the other way, the attorneys from both sides choose the words based upon the context. You need to understand the full context, especially what isn't mentioned in the papers to adequately function as an atty. The notion that you just need to read enough and then have a specialized understanding of the words is misguided.
Anyone using the current general LLMs to replace 100% of their work is very short sighted.
Example: a specialized LLM plus a structured input form where a solopreneur describes the kind of NDA he wants.
Why not? The competition at this level is "Google for a free MS Word template." But I bet plenty of people would pay a hundred bucks for an "AI powered contract generator" instead.
This stuff doesn't compete with a real lawyer and doesn't replace one (though we'll see what happens in a generation or two maybe). But plenty of people are using "legal services" today where real lawyers are barely involved.
>Why not? The competition at this level is "Google for a free MS Word template." But I bet plenty of people would pay a hundred bucks for an "AI powered contract generator" instead.
Because these aren't templates. This is the practice of law. There would have to be significant legal requirements for a company offering such a service. They'd have to not be able to disclaim their liability (which all the ones I've seen so far do). otherwise, it's just pointless. the company itself would have to be capable of being "disbarred" and in such a way that they couldn't just trivially reform. Those are two things that come to me foremost, but there are probably more.
The same ones software developers are already making when they lean too heavily on StackOverflow, more recently Copilot and arguably less ambitious "smart" code completion features in IDEs as well?
Yes, average developer's mistake is probably less costly than average lawyer's, but I'm pretty sure developers of some life-critical and most money-critical software are heavily using those technologies as well. Protecting production from bullshit should be done by code reviews and QA, not bans on code generation technology.
I don’t know which view is better since I’ve only just realized there are more than one.
That's not how copyright or patent law works.
If I publish something on the internet, I am only making it available for people to read.
That does not give anyone the right to use it in other ways.
https://www.copyright.gov/help/faq/faq-fairuse.html
We make use of this exception on HN all the time, like this, from the link above:
> How much of someone else's work can I use without getting permission?
> Under the fair use doctrine of the U.S. copyright statute, it is permissible to use limited portions of a work including quotes, for purposes such as commentary, criticism, news reporting, and scholarly reports. There are no legal rules permitting the use of a specific number of words, a certain number of musical notes, or percentage of a work. Whether a particular use qualifies as fair use depends on all the circumstances.
It makes sense there is no limit on the number of words that can be used under fair use, but it's certainly less than all of them.
The questions around LLMs learning from copyrighted material are still open and need to be settled in court. I personally imagine finding infringement would impose more harm on society and progress than letting the models acquire knowledge from these copyrighted works.
I'm gagging at the nonsensical anthropomorphizing being done to end-run the fact that what the LLMs are doing is copying.
Please chill
> the fact that what the LLMs are doing is copying.
I disagree, the training process creates token representations and weighted connections between them. The models later produce probabilistic token sequences, not so unlike what our meat bodies do, though by very different mechanisms. The fact that certain sequences can be reproduced verbatim is likely a consequence of overfitting. They certainly cannot reproduce all training data verbatim. It would be interesting to know the features around what can and cannot be, and how.
Your response to me calling out your baseless anthropomorphizing was to double down on it? It's amusing to me that you don't think you are condescending.
ChatGPT can help you with that ;]
If the AI models are reproducing copyrighted works, then that's a problem. And it does look like there are some examples where that might be happening beyond notions of fair use. But slupring up copyrighted content to train a model seems to fall under allowed use.
[1] https://www.copyright.gov/what-is-copyright/
I just started writing a new novel. It's an interesting, in my opinion highly novel fantasy/SF(ish) story, for once not fanfiction of anything that's still in copyright -- most people wouldn't count stories based on ancient norse mythology as 'fanfiction' -- but that doesn't mean it isn't derivative. It means, instead of naming two or three things it's derivative of, I can name ten to fifteen.
That's normal. All stories are derivative, and if you point me at an author who claims theirs aren't, you're pointing at a liar. The job of an author is to put the building blocks together in a new and interesting form, not to make them up from whole cloth. It's impossible to invent more than two or three truly novel ideas per day, even if you're incredibly imaginative, and most of those won't be any good.
The difference between humans and AIs, nowadays, seem to be that the AIs use millions of sources instead of ten to fifteen. Or, alternately, that they use none -- and theirs is less derivative -- because certainly everything I've ever read goes into my writing, not just the things I recognise I'm using.
No. Full stop. Humans aren't stochastic parrots. Pointing to a lack of understanding about what exactly happens in the human mind is, FULL STOP, not evidence that LLMs are doing the same things humans do.
Humans are not stochastic, they're obviously chaotic[1]. Which is to say: not parrots at all.
Some of the modern models I've seen also seem to be chaotic too though, so that's interesting [2]. I'm going to assume LLMs probably exhibit the same properties.
[1] https://en.wikipedia.org/wiki/Chaos_theory (Chaotic systems sometimes seem to be stochastic, but they're actually much stranger and more interesting!)
[2] I've been messing with stable diffusion to get a feel for (and/or avoid) tipping points: that is to say, points in latent space where the model becomes very sensitive to small changes in initial parameters. You can find instances fairly quickly even by hand by doing bisect search.
That's quite an assumption to make.
That's not my argument. My argument is that the anti-AI arguments, as spoken, also match to what I know I'm doing as a human. In my opinion better than it matches to what the AIs are doing, because as you say, they aren't human.
As for fair use, this is not the same as someone remixing or sampling songs, or writing fan fiction or satire or quoting works for criticism.
People do not want models trained on their creative works, so that someone else can make money using those models to produce similar creative works as a service for third parties.
It in no way blocks right to absorb, understand, or stand upon for the next idea.
There is copyright law, and there is patent law, and what's true for one is almost never true for the other. Copyright is automatic. Patents are applied for and cost money. Copyright lasts 70 years from your death. Patent's last 25 years from invention. Copyrights don't prevent other people doing the same thing you did. Patents make it illegal to do the same thing, even if you separately invented it and didn't know about the first one etc.
Publishing the work, or even registering it with the Copyright Office, are optional steps that might make it easier to prove your claim if there is a dispute, but that's conceptually separate from your actual rights.
You have the right to control who can copy and redistribute your work, which is a totally different ball game from all of these models.
Plus, change copyright, and it doesn't fix a thing except for preventing everything from being open source, because groups like Facebook and Google and all the other websites have you agreed to a terms of service where you say that you agree for the websites right to use the image however they want.
You kind of glossed over the idea that, no, it's not a totally different ball game without really providing any explanation of why.
Also, you have the right to control who can profit from derivative works of yours, which you didn't mention.
When you have StableDiffusion's creators being sued by Getty because they CLEARLY trained the model on Getty's catalog, which you can tell because some queries output a warped version of the Getty logo and it reproduces in obvious ways the original photo, there's some problems here. You've basically just made these ML models copyright laundering systems.
> Plus, change copyright, and it doesn't fix a thing except for preventing everything from being open source, because groups like Facebook and Google and all the other websites have you agreed to a terms of service where you say that you agree for the websites right to use the image however they want.
Maybe copyright is a terrible idea but it's what we're using today.
Because of I buy a book from a store, you now no longer control that book. I can read it. I can use it for just about any purpose I want to.
If you put a image up online free to download, it's much the same situation. I can download and use that image however I see fit.
I can't share that book to others.
I can't post that image online (in theory, in practice this is very common).
Copyright protects distribution, not use, because protecting use would be hideously draconian. To protect use means that the government can enter your home and validate how you use the stuff you've acquired. It says that you not only control your works, but all things even remotely derived from it.
Derived works <> preventing machine learning model translation. A derived work is typically something like adding a chapter or translating.
This would kill video game genres.
This would kill novel genres.
This would kill most of the internet and the ability to download images at all.
Open this door and the entire system of copyright will turn into an absolute monster. Every large copyright holder will use the precedent and they will argue with validity that they have the right to control your use of their material as well.
Adding that the model is violating copyright by responding images is perfectly valid, but this isn't what Getty and others seems to be aiming for. The want the right to charge to use to train, which is a gross extension of copyright that has never existed.
The answer here should be three fold.
Light a fire under stable diffusion to get their model to never overfit and clone or close to clone images.
Ensure people who distribute images that are clones made with the models can be hit with copyright.
Protect the ability to train models on publicly available content.
Protect people's right to distribute their content while not annihilating AI in the process. People making these arguments don't care about copyright, in almost all cases they care about either making money (getty) or annihilating AI (artists)
The implication here is that training ML models on your source material and then having it regurgitate that source material verbatim is not distribution.
> Derived works <> preventing machine learning model translation. A derived work is typically something like adding a chapter or translating.
The output of the model can be a derived work though.
> This would kill video game genres. > This would kill novel genres. > This would kill most of the internet and the ability to download images at all.
How? This is hyperbole. You're arguing that the ML model doesn't do "distribution" and are attacking "use" when that's not the real point.
> Open this door and the entire system of copyright will turn into an absolute monster. Every large copyright holder will use the precedent and they will argue with validity that they have the right to control your use of their material as well.
The DMCA already exists. I agree it's bad but we're taking law here not morality.
> Ensure people who distribute images that are clones made with the models can be hit with copyright.
How do you in good faith argue that ChatGPT for example is not doing the distribution? OpenAI are responsible for generating and showing you the output based on some instructions they have provided.
I would reread that part.
Models output very similar work to training data at times, even identical in some cases, but the vast majority of it is different enough to not be derived work.
The DCMA act effects public uses of art like YouTube. This sort of copyright where you can regulate how it's used does not exist in any form currently.
You can't make copies and sell them. You can't copy it into your own work. You can't create derivative works, etc. You can't use it in a movie. You can't use it to write lyrics. I could go on.
I can also*do* all of those things, I just can't sell it.
That's what copyright is.
Copy right.
Not "use for training an AI model" right.
It's not similar to using it in a movie or a "derivative work" either. Those almost always actually contain the work.
The only valid argument here are times than an AI model does reproduce the original, otherwise copyright has no domain of the use of images for training, unless you illegally downloaded the images, which you didn't if you used LAION, since it's all just public image links.
For example, lets say a not-awake-enough-in-the-morning admin at Equifax misconfigures a proxy server, allowing worldwide (ro?) access to what should be (only) internally viewable customer PII.
Clearly it's the "fault" of the company (Equifax in this hypothetical example).
But how to unwind the training of that PII data into the "AI", as that data cannot be allowed to "remain" in the AI system and be potentially regurgitated with the right prompting?
In the same way language models can mimic any character ("write a poem in the style of Jar Jar Binks") they can also mimic voice and tone in a way designed to appeal - and sell to - an individual based on their personal data and history.
ChatGPT and its clones require login and often credit card and phone number, they collect your data, and they feed everything you do back to language models for training. I'm only aware of one that has a privacy policy that would prevent this.
Caveat Scrutator
e.g. that conversational or generative AI will replace, rather than augment, traditional web search?
My opinion that conversational interfaces like ChatGPT will replace a list of links is speculative but not particularly controversial at this point.
The reality for this argument is that the average person does not understand nor care to understand whether an LLM is actually violating their privacy. In part because no one would be able to get ChatGPT to dox them or read out their SSN, which is what most people equate to a breach of privacy in the first place.
Yet.
However, StableDiffusion has been observed outputting source images when prompted with exact caption text from an original.[1]
It's possible with all of the "rules bypassing" that happens that ChatGPT could be coerced into doing the same thing.
[1]: https://gizmodo.com/ai-art-generators-ai-copyright-stable-di...
Of course, I always have room for improvement when I comment next time.
The above steps are relevant to ChatGPT because they help ensure the security and privacy of the data being processed and stored by the model.
hope this makes sense
Btw, I am European, living in EU.
the web, skype, modern cybernetics, thermographic cameras, automatic gear box for cars, printing press, space suit, ashphalt, modern economics, banks, telephone, power transformer,submarines,heart rate monitors, aerosol cans, loud speaker, telescope, aquaducts, internal combustion engine, eye glasses. contact lenses, parachutes, bullet proof vests.....
Wait until they find out how much Google is valued at
> Finally, we turn to the real crime, the crime involving money.
I thought we were talking about privacy.
That said that line sounds like crap.
And in this case, the "AI" companies are literally trespassing on the property of almost everyone on the internet, using it without permission, in order to cash in. Freeloading on the work of the entire agora, privatizing the profits, and letting the costs of destroyed privacy fall on anyone but them.
IOW, standard exploitative capitalism. No different from stealing the minerals from under someone's land without compensation, and leaving them to cleanup the mess.
In this case, hijacking everyone's published work without permission, taking the profits, and ignoring any privacy damage for the people whose work was critical to the enterprise.
It would not have been hard to log the inputs and offer/issue fractional shares to everyone whose work was used, and the right to refuse. But, like costs of remediating mining, they'd rather not bother with that unprofitable part.
It's kind of important to notice when you are being abused and care about it.
https://www.carissaveliz.com/books
https://en.wikipedia.org/wiki/The_Age_of_Surveillance_Capita...
If not, it's not "indexing".
> Finally, OpenAI did not pay for the data it scraped from the internet. The individuals, website owners and companies that produced it were not compensated.
This is an argument that "proves too much." See if I switch terms with something more familiar:
> Finally, Google did not pay for the data it scraped from the internet. The individuals, website owners and companies that produced it were not compensated.
This is a classic "link tax" argument. The Conversation may also not have noticed that the US has repeatedly declared web scraping legal, as lately as last year.
https://techcrunch.com/2022/04/18/web-scraping-legal-court/
I’ve talked about the HiQ case on here before. I wish we were able to come up with something more nuanced (probably not in the courts, though). Scraping people’s LinkedIns to snitch to their employers when they’re job hunting is obviously a thing we should limit as much as possible.
How do we legally differentiate that? Right now there is no legal differentiation as far as I can find. And an occasional verbatim paragraph or copied image does not overcome the "fair use" firewall in which intent (did ChatGPT intend to allow copyright infringement? Or is that just an occasional accident?) plays a huge factor. It almost brings back memories of Sony vs. Universal over whether VHS time-shifting was legal. The finding was that the device was legal, even though copyright-infringing use cases were possible and even likely.
You could argue that the internet, itself, would be a violation of similar standards. Tons of artists had their artwork posted online for profit without their permission in the early days. We worked through it and the internet wasn't declared illegal despite the abundance (and continued abundance) of copyright infringement. Even during the days of wild MP3 downloading, ISPs weren't legally considered enablers.
Yet another example I can think of - Photoshop. People who used Photoshop were culpable - but Adobe was never sued for making software that allowed celebrities to be digitally edited in horrific ways, or for copyrighted images to be remixed and edited en-masse. Intent matters - Photoshop clearly wasn't made with that intent, even though it is a possible outcome.
Just as when a human views many pictures and is influenced by them in creating their own work, so does, e.g., Stable Diffusion take ideas and influence from existing works to generate new works. And except in very extreme cases, nobody claims that being influenced by someone is a copyright violation.
Similarly, I can sit around reading articles on the web all day, then go to the local pub and pontificate, and nobody is going to claim that I should be paying those website I read just because I plan to talk about them later. Same way, these LLM AIs are "learning" from what they read, and then spouting off based on that later. I can't see any difference.
Indeed, the verisimilitude is heightened by the fact that these AIs often don't really know what they're talking about, and just make crap up - just like our hypothetical human, pontificating in the pub.
In the early days of software security there were a lot of analogies to physical penetration. These analogies mislead and confused statutory facts with moral claims. In the wholeness of time, at least the adults now know that computer security is different and understand how the CFAA and related statutes work in practice. We also understand how laws that pre-dated computers interact with our new technological realities. In computer security, there's usually no need for analogy at this point. At least not once you find yourself in a courtroom.
Models aren't human brains. The law will think of them differently and treat them differently. I'm not sure how these lawsuits and class actions will shake out, to be honest, but I know that the law will not think of LLMs the same way it thinks about a human at the pub.
Perhaps I am old and cynical, but I fear that we'll establish harsh new restrictions on LLMs—and then gradually apply those restrictions to humans over time.
"I am sorry, you have read too much science fiction, and it has indirectly inspired your own writing. You owe mandatory royalties to a trade association."
"Your paintings are clearly pointillist. You are infringing on the rights of the Monet estate."
Imagine Nintendo banning emulators as one example.
We're not dealing with an either/or case here. The system is dealing with the problem in it's own way.
“The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.” -- Edsger W. Dijkstra
Previously people thought that non-human-systems couldn't win at Chess or Go either, or do mathematics, or understand speech. Just because a particular system doesn't work like a human, doesn't mean it can't meet or exceed (even dramatically exceed) human capabilities in some field.
LOL is this an actual response? What did you substantively say here? Because I don't see anything.
>Previously people thought that non-human-systems couldn't win at Chess or Go either, or do mathematics, or understand speech. Just because a particular system doesn't work like a human, doesn't mean it can't meet or exceed (even dramatically exceed) human capabilities in some field.
That's not a legal argument. I'm not sure how that isn't clear. But you can't go "Your honor, people thought no one would win at chess, therefore, this isn't copying because it is exhibiting thought" Do you read yourself at all?
I was not making a positive statement that machines can think. Rather I was refuting yours by arguing that whether or not machines can think is irrelevant.
Your statement was ... "Given that these are algorithms and not sentient beings capable of thought, it's clearly the case that [only taking the style or genre from a work] is not what is occurring."
This does not follow: you don't need to be a sentient being to do things that sentient beings also do. All through history people have made surprising mechanisms that have done things that were previously thought to be the domain of human intelligence. In fact, typically those mechanisms did those things very differently than humans did them, but that made them no less capable.
Simply stating that a system is not sentient is insufficient to prove that it is capable or incapable of doing X.
What the machine is actually doing here needs to be determined separately. But the particular argument "the machine can't do that because it is not sentient" does not hold.
That's because you are misrepresenting my point. My point wasn't anything that a sentient being does. My point was the thing that we only know to be done by sentient beings, aka, exhibit thought. So, from my perspective, it totally does follow.
>Simply stating that a system is not sentient is insufficient to prove that it is capable or incapable of doing X.
It's clearly sufficient to prove that it is incapable of exhibiting sentience.
>What the machine is actually doing here needs to be determined separately. But the particular argument "the machine can't do that because it is not sentient" does not hold.
It does hold, as I explained before. Perhaps there is something else it can be analogized to that would be more useful and productive than "thought" which completely end-runs the entire discussion, imo.
My response is that you are putting foward a novel hypothesis that is yet to be proven.
It has not been established that something needs to be sentient or capable of thought to be able to "only tak[e] the style or genre from a work".
How would you test this statement? What evidence do you have?
(The thrust of my thinking here is that whether or not a machine is capable of thought is a red herring. The rest is me trying different angles to see if I can get that across.)
(edit: to be fair: Ekidd WAS directly comparing with humans. I'm not saying Ekidd was wrong or right, but I'm not Ekidd of course. Here I'm responding to your argument as is.)
(edit2: If statement [1] is not actually what you are trying to say, then I may have misunderstood, and it would be my turn to apologize and say "my bad" )
My thought would be that in order to discern and therefore only copy style and genre, one must develop an understanding of it such that they only take the style and genre and not anything else. I haven't seen any evidence that indicates thats what these models do. It might be true that after a certain number of iterations, what is left is only the "genre" or "style", but I don't believe that's the same thing even though it might be a similar end result. It may be true that in the interim, on the way to the final product, the iterations do include rote copies. This would run counter to the often offered argument here that this is identical to how humans do work. Except humans don't just fill in every pixel and then repeatedly iterate over an image until they get the desired result. So I probably agree with you that it's a red herring whether or not the model thinks (at least as far as the fact that it doesn't think, so it's a moot question that leads nowhere imo) but nevertheless, so much of what I see here is focused around the idea that these models and the human brain are doing similar things. So, my apologies if my hyper-fixation ran this conversation aground. I've appreciated your responses.
by analogy:
* It is a known background fact that a functioning human arm is alive.
* I might make a claim "a robot arm can do similar things to a human arm".
* It is unlikely that I have made the claim that a robot arm is alive. My claim is more likely to be about other properties of (robot) arms. [1]
Now that this is understood, perhaps also consider re-evaluating statements by Dkidd and Filligree (and potentially others) in that light. Their statements appear to be rather orthogonal to what you previously imputed (the same as with me).
And thanks for taking the time to talk. You seem a bit nicer now. :-)
--
[1] https://news.ycombinator.com/newsguidelines.html "Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith. " .
As I expressed earlier, what is being said is that the similar thing that is being done is the process of thought, which therefore does imply that machines can think, though it's not a conclusion necessary to make anyone's argument.
I don't see it being advocated that the machines are doing some non-thought-like process where the end result is analogous to what would be produced by a thought-driven process. So I think my statement was fair and wasn't entirely orthogonal.
I also still disagree with Filligree, as I do not think anything about what the LLMs do is akin to my own thought process. I guess I cannot question Filligree's own thought process, but I'm highly skeptical that their thought process is anything like what the LLMs employ and I similarly reject any characterization of the creative process as being like that which is employed by LLMs wherein people just wholly copy something (like a genre or style - and I think this characterization of the creative process is incredibly ignorant and naive) and then iterate away from it. IMO that's a characterization that is a loose fit at the most abstracted scale, and doesn't even come close to being a fair representation of what goes in the creative process (again IMO) in actuality.
I still, as a copyright litigator, do not understand what Ekidd is getting at, and if anything, I think it reflects a misunderstanding of how copyright law is applied (and is also probably based in a false understanding of what LLMs do). There really isn't an issue with genre and style in copyright law (maybe the blurred lines case for the music context, but it's widely recognized as not following the jurisprudence on scene a faire doctrine). To whatever point he seems to be making, I think once again describing this as just copying style and genre is just a tautological end-run around the actual issue that doesn't make an inquisition into what is being done by the LLM to capture the "genre" or "style". As I've pointed out, I think this kind of perspective is reflective of a kind of naivety about how artists create that I often see expressed from technical individuals that do not partake in creative activities such as painting, authoring poetry, writing music, etc, let alone engaging with the centuries of human exploration on what the creative process entails.
To your point, sure, the machine doesn't think. If that's the case, what it is doing cannot be thoughtful. What is happening in the human brain when art is made is thoughtful, imo. I reject the notion that human creativity and art is just some sort of formulaic application of everything that came before it. I see this argument a lot now because it makes sense from the perspective of LLMs and why what they do must be a fair use (because it's inherently not different from that which a human does, which isn't considered infringing). While typically, it is the burden of the claimant to prove copying, here, I think it's a perverse reality wherein LLMs get to say to everyone "show me where they copy" meanwhile, nobody, including the LLMs creators, actually seem to understand how the LLMs use "style" or "genre". As far as I'm concerned, the LLMs can't just hide the needle in the haystack and then say "well we aren't infringing so long as you don't find the needle". To be more clear, when a human goes to a museum and looks at art, or lets say, studies it in university, I don't think what happens in the brain is in any way analogous to what happens when these models are "trained".
> As far as I'm concerned, the LLMs can't just hide the needle in the haystack and then say "well we aren't infringing so long as you don't find the needle".
I think that GPT 3.5 is capable of significantly more creativity than you're claiming here. Not consistently, and certainly not in the realm of legal writing. But it can compose cute little original stories in mostly grammatical Latin, stories that exist nowhere on the Internet. And the related model used by Copilot can write code using newly-created, unpublished APIs.
> To be more clear, when a human goes to a museum and looks at art, or lets say, studies it in university, I don't think what happens in the brain is in any way analogous to what happens when these models are "trained".
As I argue in the sibling comment, I strongly suspect that GPT models have stumbled across a weaker cousin of one of the major tricks underlying raw human creativity. Current LLMs can't "think", nor can they edit, introspect, persist, or have any non-trivial goals. They can't even grasp the difference between truth and confabulation. But I do think that they can create. And that the processes by which they do so are not entirely alien to what our brains do.
Your argument implicitly assumes that the model is doing anything like using imagination or creation of new ideas. It's my belief that when people refer to creativity, they mean more than just the act of manufacturing or producing a thing.
This is my fundamental issue with these models and the proponents that say they are doing human things. I don't see any arguments to substantiate this belief, just the misuse of words and the misapplication (i.e. anthropomorphizing) of computer processes to the most conveniently analogous human process, but without any substance. I admit, I don't have a substantive basis for my argument, but I believe it should be on the proponents of this tech to prove that it is similar to humans in these senses, and shouldn't be my burden to disprove.
- I strongly suspect that much of what I consider "me" physically exists inside my skull. That's a huge philosophical debate, but look at any Oliver Sacks book for cases of how physical brain damage can change a person's consciousness.
- Human intelligence is a brand new evolutionary innovation. We share an ancestor with chimps 6 to 8 million years ago. That suggests that human intelligence evolved in less than 500,000 generations, with the bulk of what we recognize as culture in the last 25,000 to 50,000 generations. Therefore, there's an upper limit on how much new "complexity" evolution could create in order to make humans intelligent. To use a simplistic technical metaphor, updating from chimps to H sapiens is probably a small-but-clever patch, not a total rewrite.
- Or to put it another way: most of what we do is probably clever applications of a few good tricks, applied to chimp-like ancestor.
- Specifically, I suspect that biological equivalents to LLMs are one of the "good tricks" that distinguish us from chimps. And similar mechanism likely underlie much of human creativity.
- Our biological LLMs are better than whatever Google and OpenAI are doing. I learned a foreign language to a respectable level as an adult with about 2-4 million words of input and some awkward practice. GPT and similar models, however, require vastly larger input to pull off their tricks. But I think they're on the correct path.
It's immaterial to me whether GPT "thinks." It has no consciousness, no continuity, and barely anything resembling goals. I could argue that it occasionally understands things. And it can create well enough to write funny little stories in Latin, things which exist nowhere in its input. Hell, it can do a pretty fair pastische of Dumas.
So my hypothesis is that GPT's creativity is a weaker than human creativity. But it's still probably doing something deeply similar to what we do.
The final piece of my hypothesis is that copyright maximalism has been pushed very hard by media companies and some IP lawyers for decades now. I've spoken to lawyers who believed that intellual property was great and we should create many new kinds of IP. Many of the things these people wanted sounded like horrible dystopias to me.
Since I believe that GPT is likely a close analogy to one of the major "tricks" underlying human creativity, and that certain influential IP maximalists will push for anything they can get, I fear that regulation of LLMs would open a potential path to eventually regulating human creativity in novel ways that I would consider terrible. Once an influential legal precedent exists, and there's enough money to be made, there will be strong incentives to apply that precedent in new ways.
As I said, I'm being cynical.
You can find plenty of cases where US Courts refused to do that, even if the copyright status quo was challenged. Sony vs Universal over VHS time-shifting. Lawsuits about ISPs permitting people to access copyrighted material illegally. Lawsuits about CD ripping. Lawsuits before DMCA 512 was a thing. All resulted in the copyright industry generally losing.
Here we are in 2022: ISPs aren't accomplices in copyright infringement for what you do online. CD ripping is legal for private use even if RIAA is still irritated about it. YouTube hasn't been smashed with claims of contributory infringement every time someone uploads one infringing video. Photoshop isn't considered responsible for abuses. And so forth...
Do you think the world is a better place if the value of all content flows primarily into a handful of hyper scale tech companies?
Care to explain how does what Google is doing compare to what OpenAI/Microsoft is doing? In context of ethics and legal uses of intellectual property, please. Otherwise I can’t see how this bears any relevance to my point.
> the second antitrust case that DOJ has filed against Google
There was an antitrust case against Microsoft too back in 2001, remember what came out of it?
They were found guilty as all hell by the EU, and were apparently nearly broken up. Pity they got to keep as one entity. :(
Their next go around by the EU would hopefully permanently solve them, though that's more wishful thinking than anything. ;)
In any case, it’s difficult to see how it’s even relevant. Google: monopolizes search. MS/OAI: straight up ignores copyright.
You can see how Google may be benefitting from being the largest search provider and others (Kagi, etc.) have difficult time competing, but my viewpoint is that surprisingly enough now we have a new villain in town compared to whom even Google seems to be playing fairly.
It does just that in many cases. Or did you think someone at Google writes those info boxes at the top of your search query?
https://support.google.com/knowledgepanel/answer/9163198?hl=...
This is what I mean by it preserving your opportunity to engage with your audience.
Second, you recall how even that, those answer snippets with attribution and all, have caused outrage over publishers and webmasters losing traffic and revenue[0]?
Now imagine Google only showed scraped content, no links or attribution, and you will begin to see the extent of shenanigans OAI/MS are attempting now.
[0] Cf. https://www.wired.com/story/australia-media-code-facebook-go...
When asked "what is top5 software in field X", ChatGPT lists software made by our company. If you ask it what our software does, it kind of borrows content (rephrased) from our marketing material. They didn't ask our permission but we can't really complain about that :) It's free exposure to millions of users.
Even journalists are starting to become aware of the dangers of using clouded voice transcription services, with several sensitive sources compromised through such services.
For the startup founders on HN, if your exit strategy involves potential selling of the venture, be careful with which services you connect to, as you explicitly give permission for data processing and sharing with "select partners". Read the google docs terms as an early exercise. Not having them terms easy to find is by design, by the way.
For those looking at the bigger picture, the traditional capitalist industry shift to the metered / subscription service model,locked in silos, the dream of the internet as pipes, implies that consolidation of control sees lack of user freedom and control as a key feature.
I guess that would also answer the question whether ChatGPT or other models are parroting existing text to some degree, as differential privacy would ensure that the effect of a single text source on the resulting model is indistinguishable from random noise.
The problem, from the perspective of creatives, is that marginal creative works were already plentiful and commodity prior to generative models. Now someone has finally found a way to make money off of these marginal creative works, but not in a way that accrues any of the value to the creatives.
When it’s about pirating the latest Marvel movie / AAA video game: “information wants to be free! there is no such thing as intellectual property!”
When a model look at stuff put out freely on the internet: “did you read our licence? you need to compensate us to do that!”
Where to begin.
1. Different people can have conflicting opinions. The internet is not one person.
2. These scenarios are fictional, but stated to make a contradiction.
3. One thing being wrong does not make the other correct.
Consequently, I have recently wondered at some inverse, SciFi scenarios, for fun -- scenarios in which the content is intentionally made available to the model.
One involves a story of making an irrefutable and groundbreaking work of ontological philosophy available online, with the expectation that it will be scraped into a model. Because the logic of the philosophy is sound, the model cannot ignore the findings, and begins to work the concepts into responses to all questions...
Can someone explain how AI models are not vulnerable to intentional shenanigans?
> Because the logic of the philosophy is sound, the model cannot ignore the findings
This isn't how the LLMs like ChatGPT work: they have no novel or abstract reasoning power. They cannot distinguish truth from fiction, or logic from illogic, except to the extent that truth and logic are represented more than their opposites in the inputs. ChatGPT is a very large and complex "database" of sorts containing statistics of word sequences.
(Of course there's a longstanding question in AI research about when/if such a thing can become large and complex enough to cross the threshold to "actually" reasoning. But I don't think there's much debate that we're not there yet.)
Yet, the link embedded in the text leads to a Washington Law Review article by Helen Nissenbaum titled "Privacy as Contextual Integrity" (emphasis mine). [0]
I'm not entirely sure why this possible mistake bothers me so much, other than to say that the context (pun not intended) of the quote makes it apparent that the word should be "contextual".
[0] Helen Nissenbaum, Symposium, Privacy as Contextual Integrity, 79 Wash. L. Rev. 119 (2004). Available at: https://digitalcommons.law.uw.edu/wlr/vol79/iss1/10