Understood. Although personally I would file ADA case on the banning of ChatGPT and copilot as they both allow someone with a physical disability be able to work in software development.
And before you object, break your hands and try to live with speech recognition as it stands today. Then you will grab onto things like copilot as a lifesaver.
> they both allow someone with a physical disability be able to work in software development
This is somewhere between "intentionally misleading" and "objectively false" as people with physical disabilities are already able to work in software development.
> speech recognition as it stands today
...is somewhat clunky and slower than typing, but still usable. The limiting factor for software engineering is not and has never been how fast you can physically enter code.
I'm not who you're responding to, but I voice code for most of my SWE job. The technology that exists today is far better at editing existing code than writing code from scratch, making copilot and other boilerplate generators an enormous productivity gain for voice coders. The ADA is exceptionally accommodating, if the employee can make an argument that their productivity is hindered due to a disability and a tool can help, it's very likely to be approved.
"Somewhat" is such an inaccurate minimization of the impact that it borders on malicious. Voice input, when it is the only means of interacting with the computer, fucking sucks. An extremely conservative estimate would be a 10 X productivity impact for anything that requires precision keyboard input, such as programming. Most of that is due to correcting misheard terms with a side of fiddly punctuation.
To put this into context, this entire comment took me about three or four minutes to dictate. At my preinjury typing speed, it would've been done in under one, likely in under half of one.
The options for operating systems that are not Windows are limited and of typical hobbyist quality, but honestly the paid options on Windows are not that much better.
Programming does not have a 10x productivity impact by voice when using the appropriate tooling (albeit, with significant configuration), see for instance https://www.youtube.com/watch?v=uxZUJE3mC9M.
> "Somewhat" is such an inaccurate minimization of the impact that it borders on malicious.
The only maliciousness here is the extremely manipulative nature of your comment. "Somewhat clunky" is absolutely a valid way of describing the productivity impact of current voice transcription systems, and your incredible, emotionally-manipulative hyperbole is solely meant to push an agenda.
> Voice input, when it is the only means of interacting with the computer, fucking sucks.
Yes, because the current paradigm for computers (which involves either a touchscreen or keyboard and mouse, and certain ontological constructs best represented using text-adjacent symbols) is intrinsically better suited to use by people who can use their hands. You can't fix this for the current design of computers. Can it be made better? Sure. Will there ever be parity between users who can use hands and those constrained to voice? Not for computers as we know them - and Copilot won't change that.
> An extremely conservative estimate would be a 10 X productivity impact for anything that requires precision keyboard input, such as programming.
Most time spent programming is spent thinking, not typing. I do not believe your estimate.
> To put this into context, this entire comment took me about three or four minutes to dictate.
I read your comment in a normal conversational rate in under 45 seconds. If we add in another fifteen seconds to correct a few errors using a well-designed (but contemporary, non-Copilot, non-LLM) transcription tool, that's a full minute, far below your "three or four minutes" estimate - so, your complaints are about the clunkiness of current voice-transcription tools, and completely unrelated to the topic at hand, which is Copilot.
> At my preinjury typing speed, it would've been done in under one, likely in under half of one.
There are 128 words in your comment. If you typed that in one minute, you would have a rate of 128wpm, which is faster than anyone I know, and in the 99.9th percentile of typists. If you typed that in thirty seconds, that's 256 wpm. I literally do not believe you can type that fast.
You're intentionally skewing the above numbers to support your point - not that it matters anyway, because again, most time programming is not spent typing.
Barely related autoplay video that unmutes on click, only a second click stops it, can't be closed, later some completely unrelated stock data appears around the video. The people that made this website should be ashamed.
Anyways, this is ridiculous: What's their selection? I'd argue that 99% of firms world wide are not drafting policies on ChatGPT use. Gartner probably only asked HR and marketing busybodies who feel like they need a public opinion on any topic they are ever asked about to make it look like they are valuable, smart and active "thought leaders".
Until people with absolutely no notion of security and privacy start leaking API keys, internal memos, &c. to openai
It probably already happened, it happens all the time in other tools, but the scale of it and the way it's sold to the public makes it much more likely to be an issue
It's a black box, they might or might not do something now, they might or might not do it later, in the meantime the data is stored somewhere for god know how long. It might get hacked, there might be a bug allowing people to access other users history, &c.
It doesn't seem "ridiculous" to me that companies issue official statements and guidelines about what they want/not want their employees to share with a third party software. Tech illiterate people have absolutely no idea what is ok or not to share and what/where their data go.
Presumably as it’s being drafted it’s not in effect and moreover the policy may be positive for use or negative against use, but before there is one in place, it cannot apply.
My company is coming out with a policy that you can't put anything into ChatGPT. I'm ignoring it, the simple fact of the matter is it's a powerful tool. The firms that choose to use this will rapidly outpace those that don't, who will be left behind.
Maybe it's not a good idea to put your employer's source code out onto some random website. If you wouldn't put it on here, you probably shouldn't put it on ChatGPT. I'm sure it's only a matter of time there will be a code ai that can be run behind a firewall but until then stop giving away the secret sauce.
Firms that choose to use it indiscriminately will also be exporting domain knowledge (likely including those that give them a competitive advantage) into ChatGPT.
This doesn't matter in roles that are complementary to where the most value is generated (the phrase "commoditize your complement" exists for a reason) but that's not all roles.
Congratulating, you're sending your work products to a third party under extremely permissive agreements. You may be:
* Rendering invalid any attempts to assert intellectual property (trade secrets, patents).
* Destroying attorney-client privilege
* Violating consumer privacy laws
* In some other sectors, you may have even stronger no-data-sharing laws that get violated (e.g., HIPAA, FERPA).
I suspect your company's policy is based on legal concerns such as the above... do you really understand the legal risks of violating your company's policy? Are you really willing to take on all that legal risk personally (the company sure isn't going to protect your ass)?
You need to think about where you're going before you just blindly follow "progress". You're assuming that you'll be better off than the ones left behind, I wouldn't always be so sure.
You are aware that openai explicitly says they send the prompts and answers to affiliated parties that read them and rate them to put them back into their reinforcement learning database?
Has OpenAI said anything (in their TOS/etc) about who owns the copyright/ip on the code that it generates?
I know it’s a legal minefield of a question, just curious if they have said “we won’t sue you if you copy/paste this code in your app” publically or anything.
That might be the case in theory, but if a university can't tell the difference between AI generated essays/books and human generated ones the how could the courts?
If someone tries to sue you for copyright infringement on a model output, you can just produce the parameters to the model that produced the output and the case will be thrown out pretty much instantly.
This feels like it could easily be abused if you can get a model to output code that is "pretty close". You could use it to try and invalidate human written copyright as well.
What happens when you incidentally created something very similar to something ChatGPT could produce - does that mean you can't enforce copyright on it? (as someone could automate a prompt engineering process to create a near-copy of your original work?
Let's take that to the next step - what happens if ChatGPT happens to get trained on something that is copyrightable - then a prompt script can be used to reproduce the original - does that mean the original is now considered not protected by copyright?
This is a lot more murky than "show me the prompts/parameters" and it's not protected by copyright.
If you try to sue someone for copyright infringement and they demonstrate that what you're claiming is infringement is solely the result of model output, you're going to be out of luck, and honestly I don't think that's a problem - if your creation is so derivative that a model produces it via a reasonable prompt it probably doesn't deserve protection. Using prompt interrogation on a specially fine tuned model in an attempt to end-run a reasonable copyright is a different matter, but I don't think that'd be worth the effort for anyone to actually do.
If a copyright precedes the existence of a model, the model was trained on that code and someone distributes code from the model that is a near-verbatim reproduction, that's a different story (in large part because getting that near verbatim reproduction would probably require willful intent rather than just being accidental)
Can you explain to a court why it is “generating” the original from scratch from a prompt and not simply retrieving the input and printing it to screen?
Just out of curiosity.
I don’t think the owner of the copyright loses copyright simply because something else output a copy of his work. Maybe it is “generated” but the copyright holder still has his rights per copyright statutes and you would still be violating those rights by distributing your own copies from the AI output.
As anyone well-informed how LLMs work will tell you, it's not violating any copyright or IP, because these answers are coming from a probabilistic walk through condensed remix of learning, not the sources themselves.
Another way of thinking of it is compressed knowledge, with learned structure, and an ability to pick likely words to describe that structure rather than an ability to use original words.
The compression ratio of the source corpus versus the model demonstrates it doesn't have all the source material, ipso facto, results are not infringement.
As an easy to understand analog, if you grab a Shutterstock photo, change it from 100% quality to 5% quality, the file size drops to 100th of the original and on display, none of the pixels are the same. While it may be the same general concept depicted badly, it is certainly not the original and clearly isn't copyrightable.
(Note: That last paragraph is firmly tongue in cheek /s ... I've used the "uniqueness (originality) is compressed out, what remains as learned is fair use" argument in the past, but on further reflection about the interaction of creativity and structure, I'm not sure it stands, especially at temperature 0 aka zero creativity.)
It should be noted that the GP was asking a different question - they were asking whether OpenAI might have any claims of copyright over the output of their model, so that they might sue you for using the output of their model for your own business.
Hmm.. so there is also procedural computer art, writing code to produce nice images. Also not copyrightable? At least the code?
I draw an image.
I write some lines of code that draws an image.
I write many lines of code that do something, gather something, and then draws an image?
I just wonder how one can reasonably draw hard lines (and now I don't mean art :D) in the future.. there are so many ridiculous incarnations of copyright and patent fights already out there..
The principles are quite clear: Copyrightability is based on how much originality of a work is attributable to a human (because copyright law is limited to human creations). See [0]. In case of dispute, this has alway had to be assessed by a court on a case-by-case basis, there were never hard lines.
In the context of using AI or coding, the question is how much creative control you have over the output. When writing code to draw an image pixel by pixel, then you have full control, probably even for procedural art when you devised the algorithm completely on your own. When using an existing library that can draw crayon strokes or produce nice water color fills, there is less attributable to your own creativity. Similar for AI, where it will have to be assessed how much of the creativity is attributable to the human-invented prompting (likely not much) vs. to the AI.
Humans can claim processes as intellectual property though. The question in my mind is: if I develop AI that to perform some specific task (like image generation), does that constitute a specific patentable process? And thus, if you use that process without compensation, are you stealing my intellectual property?
Shouldn't this depend on the licensing terms of the AI you are using to generate the content. I haven't read the various terms of current offerings. But it is probably not unreasonable to include bad terms in there...
Maybe, assuming you are using a third-party's tool directly. But I guess I was thinking more about the issue of someone creating a process that is more easily replicable.
There are lots of instances of fairly fragile architectures that work really well in a particular application, but if you change some small part of that architecture, they fail. Now say I spend a lot of research time identifying that architecture and maybe even patent it as a process. I could even publish the methods of exactly what that process entails and somebody could copy it an use it. But I think that would be illegal (assuming AI can be patented in this way).
Fair use is a judgement. Fair use is an affirmative defense (admitting guilt, but saying you should be allowed to, like self defense.) A judge looks at the work and the framework and decides if fair use is an acceptable defense to infringement.
Given most of what we are talking about would fall under the output being used commercially, lets look at the end of of fair use test.
> the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
>the effect of the use upon the potential market for or value of the copyrighted work.
Unless your opponent is using a model exclusively trained on Twilight to output a Twilight sequel before you the author, you are going to have a have a hard time saying they used a large part of your work and that it directly displaces a sale of your work.
A collage of word samples from millions of works is going to fall under fair use. If the model did a good enough job of compressing your work in its entirety, and it outputs that copy, you might have a case.
The OP was asking a different question with a different answer -- OP wanted to know if novel ChatGPT output is copyrightable, not whether existing copyrights might cover ChatGPT outputs. That said...
> The compression ratio of the source corpus versus the model demonstrates it doesn't have all the source material, ipso facto, results are not infringement.
What?! This is not at all how copyright law works and that's also not at all how LLMs or compression work! You're wrong about both the law and the technology!
Just because the model can't losslessly reproduce the entire training dataset doesn't mean that the model can't losslessly reproduce copyrightable fragments of the training dataset.
And the law truly doesn't give a damn that a verbatim copy of a piece of code or sequence of paragraphs came out of a model instead of a copy/paste from github. If it's the same basket of bits, or close enough in certain cases, it's covered by copyright, full stop.
> As an easy to understand analog, if you grab a Shutterstock photo, change it from 100% quality to 5% quality, the file size drops to 100th of the original and on display, none of the pixels are the same. While it may be the same general concept depicted badly, it is certainly not the original and clearly isn't copyrightable.
As an easier to understand analog, if you grab a BITMAP of a Shutterstock photo and convert the file type to a reasonable JPEG (lossy compression!!!), tht's very clearly still covered by the copyright on the .bmp.
> The compression ratio of the source corpus versus the model demonstrates it doesn’t have all the source material, ipso facto, results are not infringement.
Or, why making and redistributing a JPEG of a copyright-protected artwork is not infringement.
Results decidedly not guaranteed if you try this “a copy with lossy compression is not infringement” argument in an actual court of law.
"OpenAI hereby assigns to you all its right, title and interest in and to Output. This means you can use Content for any purpose, including commercial purposes such as sale or publication, if you comply with these Terms."
The company I work for, literally blocked the OpenAI site at the network level "until they can assess the potential security impact". I guess they fear people might leak sensitive information to OpenAI. Very frustrating for people with a legitimate use case for it.
But Dropbox and Google are not blocked, people are just told not to upload random company data to it. Same could be done with OpenAI with a simple email.
My company has a group of younger product managers who are very obviously using ChatGPT to write proposals, e-mails, and other documents.
It's obvious because their writings in Confluence and late night e-mails suddenly became very verbose and used the sometimes obvious ChatGPT voice. There's a stark contrast between content they write/share on the spot or discuss in meetings.
Then you get them into a meeting to explain further and they some times can't even remember what they were thinking when they wrote various bullet points.
It's eye opening to see how much the ChatGPT produced content can fool managers at the top who catch glimpses of neatly composed and verbose documents. Then you get the people into meetings and they don't really know where they were going with their "own" documents.
It wouldn't be a problem if they could take the documents as some sort of next step for their learning or research, but it's really becoming useless to have all of this ChatGPT-created content out there where none of them can really execute on it without going back to ChatGPT over and over again to come up with more content.
> It's eye opening to see how much the ChatGPT produced content can fool managers at the top who catch glimpses of neatly composed and verbose documents.
Goodhart's law in action. If "neatly composed and verbose documents" is the metric higher ups use to judge product managers, then documents will become neater and more verbose to the detriment of everything else.
In the end, the alignment problems that plague AI safety research are all playing out in company structures every day. It's hard to give a metric that measures your intended outcome, and any proxy leads to unintended consequences once there's an easier way to reach that proxy.
In the end, the alignment problems that plague AI safety research are all playing out in company structures every day. It's hard to give a metric that measures your intended outcome, and any proxy leads to unintended consequences once there's an easier way to reach that proxy.
I've actually pointed this out as a fallacy of the premise of current alignment theory. That being we simply need to apply humanities values for the AI to be properly aligned. However, every alignment problem can also be seen as an existing human problem for which we have never resolved.
The scale is the key. Things that have impact at scale are harmless initially and then in the long run the damage becomes visible and by then we'll be as dependent on them as a heroin junk is on their fix.
Oil, plastic, electricity, natural gas, mines, wood, poultry, meat & fish, transportation etc. They're nice and super useful and when used in moderation or when there aren't that many humans to begin with and you can easily absorb the consequences. But if you multiply the number of people while at the same time you increase the consumption/production on some or even all of these in tandem the effects are massively compounding.
Yes, but that is just one facet. Alignment theory affects everything regardless of scale. It is about whether or not you can reliably get the intended result given a particular goal or task.
We want AI to reason like humans and have humanities values. According alignment theory. Yet humans will also cheat to get a result. Our values also are what put us in conflict with each other.
>the alignment problems that plague AI safety research are all playing out in company structures every day.
Yes yes yes holy crap yes.
Corporations are AI, and we are already in the AI apocalypse. Corporations exhibit emergent behavior that any individual human involved with them is powerless to stop, and they self-optimize towards inhuman goals. Coca Cola inc is a paperclip maximizer, except it maximizes the consumption of flavored sugar water. We are completely failing as a society to adequately rein them in.
There is this movie called Idiocracy. When I saw it the first time I thought it was hilarious. But over time I've stopped laughing and I now more or less realize that the makers may well have intended something funny but accidentally ended up making a prophesy. Or maybe they intended that from the get-go but they dressed it up as something funny for budget reasons.
It's scary how many of the things that are in that movie as caricatures you can now find IRL.
Mike Judge is a genius. He also did Office Space and Silicon Valley, among other things (I personally love King of the Hill). He has a sharp eye for both society and humor.
> The problem isn't thinking machines, it's letting machines do the thinking for you
...and then I like to point out that many of our social structures (including corporations) are, essentially, machines. They're just machines running on meat instead of silicon.
> There's a stark contrast between content they write/share on the spot or discuss in meetings.
I'd be worried about false accusations of this. Some people are much better at communicating in writing, when they have time to think about how to word things, and not so good when put on the spot in a meeting. Others might be better at communicating verbally, and perhaps too verbose in written communication.
Shouldn’t fire off on someone for seeming smarter in writing than in meetings. But if they don’t even seem to understand the things “they” wrote about and they finish JIRA tickets with “in conclusion”…
You're exactly right: This isn't a situation to jump to conclusions.
However, when the junior product managers are proposing things with terms they don't even understand then something is clearly wrong. Imagine joining a meeting where a product manager is trying to explain their own document but can't understand what "they" wrote when they get to various bullet points. Or the document is correct, but their understanding of their own document is incorrect.
Have you ever seen a student caught cheating? Someone who has the right answers on the paper, but can't explain why or remember why they wrote it down? It's exactly that.
I'm working on an open source startup to allow companies to self host OSS AI/LLM models on their own infrastructure, whether through AWS/GCP/Azure/DigitalOcean/etc or on their physical servers. I'm fine tuning models with prompts and then I'll build out the API so that users and other developers can query them. This way, you send nothing back to a proprietary service.
Absolutely. I'm just an IC though, so I have no sway. That being said, this is something I've been talking about for a while now with co-workers. Being able to run an LLM on your codebase and query it just for your code would be fantastic!
I would expect most firms to have policies in place prohibiting employees to share work-related information with outside parties they don’t have a pertinent business relationship with. In principle that already covers (and rules out) most uses of ChatGPT for work.
88 comments
[ 2.8 ms ] story [ 152 ms ] threadIn other news - Nearly half of firms doing nothing about ChatGPT use.
And before you object, break your hands and try to live with speech recognition as it stands today. Then you will grab onto things like copilot as a lifesaver.
It's the same rhetoric they use to try to push for backdoor in OSes, messaging services, &c. "but what about the pedophiles, think of the children!"
That is a LOT to ask me to do before objecting. I'm pretty good at imagination, so I'll probably skip that step.
This is somewhere between "intentionally misleading" and "objectively false" as people with physical disabilities are already able to work in software development.
> speech recognition as it stands today
...is somewhat clunky and slower than typing, but still usable. The limiting factor for software engineering is not and has never been how fast you can physically enter code.
To put this into context, this entire comment took me about three or four minutes to dictate. At my preinjury typing speed, it would've been done in under one, likely in under half of one.
The options for operating systems that are not Windows are limited and of typical hobbyist quality, but honestly the paid options on Windows are not that much better.
The only maliciousness here is the extremely manipulative nature of your comment. "Somewhat clunky" is absolutely a valid way of describing the productivity impact of current voice transcription systems, and your incredible, emotionally-manipulative hyperbole is solely meant to push an agenda.
> Voice input, when it is the only means of interacting with the computer, fucking sucks.
Yes, because the current paradigm for computers (which involves either a touchscreen or keyboard and mouse, and certain ontological constructs best represented using text-adjacent symbols) is intrinsically better suited to use by people who can use their hands. You can't fix this for the current design of computers. Can it be made better? Sure. Will there ever be parity between users who can use hands and those constrained to voice? Not for computers as we know them - and Copilot won't change that.
> An extremely conservative estimate would be a 10 X productivity impact for anything that requires precision keyboard input, such as programming.
Most time spent programming is spent thinking, not typing. I do not believe your estimate.
> To put this into context, this entire comment took me about three or four minutes to dictate.
I read your comment in a normal conversational rate in under 45 seconds. If we add in another fifteen seconds to correct a few errors using a well-designed (but contemporary, non-Copilot, non-LLM) transcription tool, that's a full minute, far below your "three or four minutes" estimate - so, your complaints are about the clunkiness of current voice-transcription tools, and completely unrelated to the topic at hand, which is Copilot.
> At my preinjury typing speed, it would've been done in under one, likely in under half of one.
There are 128 words in your comment. If you typed that in one minute, you would have a rate of 128wpm, which is faster than anyone I know, and in the 99.9th percentile of typists. If you typed that in thirty seconds, that's 256 wpm. I literally do not believe you can type that fast.
You're intentionally skewing the above numbers to support your point - not that it matters anyway, because again, most time programming is not spent typing.
It’s absolutely the right thing to consider, just as social media rose organizations drafter social media policies.
Anyways, this is ridiculous: What's their selection? I'd argue that 99% of firms world wide are not drafting policies on ChatGPT use. Gartner probably only asked HR and marketing busybodies who feel like they need a public opinion on any topic they are ever asked about to make it look like they are valuable, smart and active "thought leaders".
Until people with absolutely no notion of security and privacy start leaking API keys, internal memos, &c. to openai
It probably already happened, it happens all the time in other tools, but the scale of it and the way it's sold to the public makes it much more likely to be an issue
OpenAI aren't training new models on raw input to their bots (they're using it for other levels of improvement).
Are you worried that an OpenAI employee might steal API keys out of their logs?
(I still wouldn't paste an API key in myself.)
It doesn't seem "ridiculous" to me that companies issue official statements and guidelines about what they want/not want their employees to share with a third party software. Tech illiterate people have absolutely no idea what is ok or not to share and what/where their data go.
Best not to leave it to the users to try to figure out which ones do and don't. They already post secrets to StackOverflow and Github.
Not all about API keys either. Think of it like the new confession booth, and AI is the new priest. Lots of valuable intelligence is run through them.
It's too soon to see these boundaries fail, but they're supposed to be the "safe" option for this.
I also can’t paste anything in ChatGPT window, there are at least 3 spying/tracking software pieces on my work computers. I will comply.
This doesn't matter in roles that are complementary to where the most value is generated (the phrase "commoditize your complement" exists for a reason) but that's not all roles.
* Rendering invalid any attempts to assert intellectual property (trade secrets, patents).
* Destroying attorney-client privilege
* Violating consumer privacy laws
* In some other sectors, you may have even stronger no-data-sharing laws that get violated (e.g., HIPAA, FERPA).
I suspect your company's policy is based on legal concerns such as the above... do you really understand the legal risks of violating your company's policy? Are you really willing to take on all that legal risk personally (the company sure isn't going to protect your ass)?
You need to think about where you're going before you just blindly follow "progress". You're assuming that you'll be better off than the ones left behind, I wouldn't always be so sure.
Let's take that to the next step - what happens if ChatGPT happens to get trained on something that is copyrightable - then a prompt script can be used to reproduce the original - does that mean the original is now considered not protected by copyright?
This is a lot more murky than "show me the prompts/parameters" and it's not protected by copyright.
If a copyright precedes the existence of a model, the model was trained on that code and someone distributes code from the model that is a near-verbatim reproduction, that's a different story (in large part because getting that near verbatim reproduction would probably require willful intent rather than just being accidental)
Just out of curiosity.
I don’t think the owner of the copyright loses copyright simply because something else output a copy of his work. Maybe it is “generated” but the copyright holder still has his rights per copyright statutes and you would still be violating those rights by distributing your own copies from the AI output.
Another way of thinking of it is compressed knowledge, with learned structure, and an ability to pick likely words to describe that structure rather than an ability to use original words.
The compression ratio of the source corpus versus the model demonstrates it doesn't have all the source material, ipso facto, results are not infringement.
As an easy to understand analog, if you grab a Shutterstock photo, change it from 100% quality to 5% quality, the file size drops to 100th of the original and on display, none of the pixels are the same. While it may be the same general concept depicted badly, it is certainly not the original and clearly isn't copyrightable.
(Note: That last paragraph is firmly tongue in cheek /s ... I've used the "uniqueness (originality) is compressed out, what remains as learned is fair use" argument in the past, but on further reflection about the interaction of creativity and structure, I'm not sure it stands, especially at temperature 0 aka zero creativity.)
If a human didn't create it, it can't have a copyright.
I draw an image. I write some lines of code that draws an image. I write many lines of code that do something, gather something, and then draws an image?
I just wonder how one can reasonably draw hard lines (and now I don't mean art :D) in the future.. there are so many ridiculous incarnations of copyright and patent fights already out there..
In the context of using AI or coding, the question is how much creative control you have over the output. When writing code to draw an image pixel by pixel, then you have full control, probably even for procedural art when you devised the algorithm completely on your own. When using an existing library that can draw crayon strokes or produce nice water color fills, there is less attributable to your own creativity. Similar for AI, where it will have to be assessed how much of the creativity is attributable to the human-invented prompting (likely not much) vs. to the AI.
[0] https://en.wikipedia.org/wiki/Threshold_of_originality
There are lots of instances of fairly fragile architectures that work really well in a particular application, but if you change some small part of that architecture, they fail. Now say I spend a lot of research time identifying that architecture and maybe even patent it as a process. I could even publish the methods of exactly what that process entails and somebody could copy it an use it. But I think that would be illegal (assuming AI can be patented in this way).
Training an ML model doesn't seem to fall under the established exceptions which qualify as "fair use".
Given most of what we are talking about would fall under the output being used commercially, lets look at the end of of fair use test.
> the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
>the effect of the use upon the potential market for or value of the copyrighted work.
Unless your opponent is using a model exclusively trained on Twilight to output a Twilight sequel before you the author, you are going to have a have a hard time saying they used a large part of your work and that it directly displaces a sale of your work.
A collage of word samples from millions of works is going to fall under fair use. If the model did a good enough job of compressing your work in its entirety, and it outputs that copy, you might have a case.
> The compression ratio of the source corpus versus the model demonstrates it doesn't have all the source material, ipso facto, results are not infringement.
What?! This is not at all how copyright law works and that's also not at all how LLMs or compression work! You're wrong about both the law and the technology!
Just because the model can't losslessly reproduce the entire training dataset doesn't mean that the model can't losslessly reproduce copyrightable fragments of the training dataset.
And the law truly doesn't give a damn that a verbatim copy of a piece of code or sequence of paragraphs came out of a model instead of a copy/paste from github. If it's the same basket of bits, or close enough in certain cases, it's covered by copyright, full stop.
> As an easy to understand analog, if you grab a Shutterstock photo, change it from 100% quality to 5% quality, the file size drops to 100th of the original and on display, none of the pixels are the same. While it may be the same general concept depicted badly, it is certainly not the original and clearly isn't copyrightable.
As an easier to understand analog, if you grab a BITMAP of a Shutterstock photo and convert the file type to a reasonable JPEG (lossy compression!!!), tht's very clearly still covered by the copyright on the .bmp.
Firmly tongue in cheek sarcasm, as in, this is obviously wrong.
Or, why making and redistributing a JPEG of a copyright-protected artwork is not infringement.
Results decidedly not guaranteed if you try this “a copy with lossy compression is not infringement” argument in an actual court of law.
You've replied in the spirit intended. :-)
OpenAI ToS 3(a):
"OpenAI hereby assigns to you all its right, title and interest in and to Output. This means you can use Content for any purpose, including commercial purposes such as sale or publication, if you comply with these Terms."
It's obvious because their writings in Confluence and late night e-mails suddenly became very verbose and used the sometimes obvious ChatGPT voice. There's a stark contrast between content they write/share on the spot or discuss in meetings.
Then you get them into a meeting to explain further and they some times can't even remember what they were thinking when they wrote various bullet points.
It's eye opening to see how much the ChatGPT produced content can fool managers at the top who catch glimpses of neatly composed and verbose documents. Then you get the people into meetings and they don't really know where they were going with their "own" documents.
It wouldn't be a problem if they could take the documents as some sort of next step for their learning or research, but it's really becoming useless to have all of this ChatGPT-created content out there where none of them can really execute on it without going back to ChatGPT over and over again to come up with more content.
Goodhart's law in action. If "neatly composed and verbose documents" is the metric higher ups use to judge product managers, then documents will become neater and more verbose to the detriment of everything else.
In the end, the alignment problems that plague AI safety research are all playing out in company structures every day. It's hard to give a metric that measures your intended outcome, and any proxy leads to unintended consequences once there's an easier way to reach that proxy.
I've actually pointed this out as a fallacy of the premise of current alignment theory. That being we simply need to apply humanities values for the AI to be properly aligned. However, every alignment problem can also be seen as an existing human problem for which we have never resolved.
Oil, plastic, electricity, natural gas, mines, wood, poultry, meat & fish, transportation etc. They're nice and super useful and when used in moderation or when there aren't that many humans to begin with and you can easily absorb the consequences. But if you multiply the number of people while at the same time you increase the consumption/production on some or even all of these in tandem the effects are massively compounding.
We want AI to reason like humans and have humanities values. According alignment theory. Yet humans will also cheat to get a result. Our values also are what put us in conflict with each other.
FYI, I've elaborated in much more detail here on such issues with containment/alignment - https://dakara.substack.com/p/ai-singularity-the-hubris-trap
Yes yes yes holy crap yes.
Corporations are AI, and we are already in the AI apocalypse. Corporations exhibit emergent behavior that any individual human involved with them is powerless to stop, and they self-optimize towards inhuman goals. Coca Cola inc is a paperclip maximizer, except it maximizes the consumption of flavored sugar water. We are completely failing as a society to adequately rein them in.
It's scary how many of the things that are in that movie as caricatures you can now find IRL.
> The problem isn't thinking machines, it's letting machines do the thinking for you
...and then I like to point out that many of our social structures (including corporations) are, essentially, machines. They're just machines running on meat instead of silicon.
I'd be worried about false accusations of this. Some people are much better at communicating in writing, when they have time to think about how to word things, and not so good when put on the spot in a meeting. Others might be better at communicating verbally, and perhaps too verbose in written communication.
However, when the junior product managers are proposing things with terms they don't even understand then something is clearly wrong. Imagine joining a meeting where a product manager is trying to explain their own document but can't understand what "they" wrote when they get to various bullet points. Or the document is correct, but their understanding of their own document is incorrect.
Have you ever seen a student caught cheating? Someone who has the right answers on the paper, but can't explain why or remember why they wrote it down? It's exactly that.
Is anyone else interested in something like this?
Albeit, it is using proprietary GPT instead.