This reminds me of a billboard I saw about 15 years ago when the state of Delaware was considering changing its laws to permit new casinos in addition to the three that had been grandfathered in when the restrictions went into place.
Delaware Park, one of those three casinos, put up a billboard opposing the bill, and it said, "No more casinos! THREE IS ENOUGH!"
It just goes to show that corporations will take any position -- even a casino taking an anti-casino position -- as long as it's self-serving.
> The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
The New York Times lawsuit isn't purely about training on content - it's about memorization. They included examples of OpenAI's models being able to reproduce NYT articles entirely from the weights.
There's also stuff about RAG in there - about OpenAI (and Microsoft) features that run web searches and then summarize the results. I've not been following developments closely, I'd be interested to know how that part of the lawsuit is shaking out so far.
Right, it's not just that they saw the thing but also how they use the knowledge later.
To demonstrate this "what you do next matters" part, assume I get a book from the library and I read it, and I have perfect memory. So far, no problem, even with my fantastical memory. However what I do next matters:
A. I quote several short pieces of the book in a written critique. Fair use.
B. I use what I know to create a statistical breakdown of the author's favorite words and phrases. That's fine too.
C. I retype my own book which is 90% identical to the original and I publish it for money. Nuh-uh. Legal problems abound.
We can debate whether or not these companies are crossing the line, but the line does depend on what they do with the information, not just how they got it.
The reproduction argument was always the weaker of the arguments. The whole point of news is that it's new, if you ask ChatGPT for the current front page story that's not gonna be in the weights. They were able to contrive a single example of a very well known piece, but the number of people who will use that to circumvent the paywall is zero.
Honestly if you are putting models behind your own terms of use it's only fair use if I get to do the same for my content, but it's too late for me I folks in US sue them for every dollar or force them to make their models "fair-use".
I don't think scooping all of the internet and saying rules for thee but not for me is the way to go.
Exactly, the same copyright rules should and do apply to both parties.
I'm glad copyright law didn't prevent the creation of LLMs. I'm also glad it doesn't prevent others from creating LLMs. I think that benefits us all.
Without government intervention, I think distillation will prevent these giant AI companies from recouping their investment. But that's not our problem.
The other LLM is obviously transformative because it's a different LLM and therefore is not even the same kind of thing, like how a white horse is not a horse, obviously.
The key point in all these AI training lawsuits is "transformative use".
If an LLM trains on NYT articles is the output of the model a substitute for the original article? Or will people continue to pay for the New York Times as before?
Whether they build a business around it doesn't really matter. Plenty of large businesses depend on fair use to avoid copyright lawsuits, including any tech company that hosts any kind of content (YouTube, Reddit, Meta). Even the media itself relies on fair use to be able to air any content.
> Plenty of large businesses depend on fair use to avoid copyright lawsuits, including any tech company that hosts any kind of content (YouTube, Reddit, Meta).
Those companies don't rely on fair use. They rely on the DMCA. There's an agreement with the user for use of content where the user controls the rights, and anything that users upload where someone else controls the rights, DMCA protects the site as long as they follow the procedures.
> Does building a business around the automation of fair use constitute fair use? I would say no.
This seems like an excellent and very clear distinction. Once it becomes a wholesale harvesting and processing for financial gain, it stops being fair use. Fair use clearly has a magnitude component.
The irony of Sam Altman's company pretending to care about what's "fair" is so very rich.
He literally made billions exploiting the intellectual work of millions of his fellow humans, yet has the audacity to whine about how the system isn't fair.
Generational level of hypocrisy from OpenAI. What a nerve putting "authors" in the list first. Fair Use benefits OpenAI and the labs much more than anyone else.
Just be honest and don't try to blend it into "Everyone".
Fruitfulness of discussion might be aided by bearing in mind that US and European copyright law traditionally come from very different perspectives. Basically, societal utilitarianism vs work-is-an-inalienable-part-of-the-artist author's natural rights. "Copyright should last 5 years because starving keeps artists productive. And copyright owners have no say at all in their works being used for LLM training because it's good for society" is consistent with US foundations, at least pre-Berne. "Authors are not permitted to give up their right to personally refuse a use of their already-sold work, nor to give anticipatory permission for use in future technology" is IIUC unremarkable in a European context. An AI suggests[1] as philosophical background. People are coming at this from vastly different places. Bearing that in mind might help avoid talking past one another. IIUC.
I also note, when the internet was just catching on, and email was becoming less niche, but copyright law had not yet established sending email as an implicit grant of permission to make copies, most all use of email violated copyright. Network->disk->memory - all copies without permission. I'd get email with footers vaguely like "You are hereby granted permission to <do email> for the purposes of XYZ Corp - all other use constitutes infringement". There was a multi-year lag between accepted practice and legality. IIUC, forwarding email is still problematic today. Having lived through that, "Arg! <New tech> violates copyright!" gets less traction with me. And with copyright term extensions, and Berne on the US side, having a taint of unprincipled corruption, that also makes it less compelling.
But asking the trained model lots of questions, and creating a similar but different open weights model? No, that's fair use.
The current copyright rules cover both cases well and benefit everyone, I think. (I think copyrights are waaay too long, but that's another topic....) Whether trillion dollar companies can ever make a profit once open weights competitors distill them is, frankly, not our concern.
> Authors, journalists, creators, academics, accessibility advocates, legal experts, librarians, economists and business leaders are voicing their support for the longstanding principle of fair use and its application to AI training.
Have you heard of the landmark $1.5 billion class-action copyright settlement in Bartz v. Anthropic?
The Bartz v. Anthropic settlement wasn't about fair use, it was about pirating ebooks. The judge in that case said that the training on scanned, purchased books was fair use - Anthropic started doing that in 2024.
And it supports my summary here. Anthropic did "win" on the fair use issue, but lost on the pirating.
Note that I said "The Bartz v. Anthropic settlement wasn't about fair use" - and that's accurate, the settlement aka the $1.5bn was about the pirated books, even though the initial lawsuit was about fair use.
ironic that they can talk out of both sides of their mouth, fair use is important for their training but if you distill it's definitely not fair use anymore
Copyright originally maxed out at 28 years. Corporations rewrote the laws over and over, so that now copyright can last over 100 years, turning intellectual property into yet another rent extraction scheme.
Now, in this case, OpenAI argues that “fair use” is for the good of all. OpenAI doesn’t give a shit about what’s good for all. They want the ability to define intellectual property rules so that they can extract rent from someone else’s property.
If they want to benefit everyone, why aren’t their models open? For that matter, why is their name OpenAI, when they are not open? They’re publishing tons of work in mathematics, supposedly to benefit everyone. Instead, why don’t they forgo all research themselves and dedicate their work to supporting existing mathematicians in using AI effectively, so that the people that built the base of knowledge get credit for the new work?
Because OpenAI wants to extract the maximum shareholder value out of other people’s work.
Does anyone actually believe PR like this anymore?
I'm not sure what I find more morally repugnant to be honest. The largest wholesale theft of human labour ever seen, or the company that has made billions from this heinous act disingenuously and cynically trying to claim it as in any way "fair" (whilst simultaneously crying foul about distillation)
This is a wild re-imagining of what "Fair use" means in the US courts, which have a relatively straightforward four point test, and of which OpenAI, a commercial venture that produces exact copies of an article and thus acts as a market-replacement for the original work, fails handily.
What AI is doing is socializing the costs but privatizing the profits. This has been the model for every social media platform going back to forums. Everything after USENET, really.
If you're using the collective output of humanity to train your LLM then humanity should own the output. Problem solved.
One can easily understand this as the worker's relationship to the means of production.
But you haven't. You've had maybe 40 years to read books, web pages, watch movies listen to music etc. Every LLM has millions of books, nearly every web page ever written, and every image available on the web. There's not even a rough order of magnitude comparison, your content consumption is not even a rounding error on the amount of text every LLM has in its weights.
I've had that same thought. It's a surprisingly valid point even though people tend to think you're joking. A human is capable of consuming a certain percentage of all human output, admittedly a low percentage, but that might change with, say, vastly extended lifespans in the future. What percentage is allowed?
The first difference is that somebody is typically getting paid for any content you consume. You buy books, TV shows, movies, magazines, whatever. Or you borrow them from a library who bought them. Or you pay a subscription for a streaming service. Or you use an ad-supported platform. If you bought any commercial IP and then reproduced it one million times so you can train a whole bunch of people at once, well that's a crime. So why isn't it a crime when AI companies do it?
The second is that you can only produce so much output. You can't be cloned a million times to produce a million times the output. Software can be.
LLMs are kind of like a billion people sharing one NYT account, 1 Netflix account and so on. And that's simply not sustainable.
I also believe that the lifespan of IP is way too long. Software shouldn't be patentable because it's ultimately math and math formulae aren't patentable. Other patents should generally have much shorter lives. Books, music, TV, movies should have much shorter copyright periods, say 10-30 years total.
AI companies ingesting all the IP on earth for free is like saying there is no IP at all. Some will agree with that. I don't. But again, I would have less of an issue with it if we as a society owned the output.
> The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
There's absolutely no "fair use" precedent that says that you can program a computer with copyrighted material. They're just asking for an exception essentially based on their compression being lossy, which of course the law takes the absolute opposite opinion on e.g. distributing a 1080p of my bluray disc has been determined to definitely not be fair use.
Their essential case is that they use the word "training," and people training themselves from books is fair use, so since they use the word "training" then what they're doing is fair use. It's stupid, and judges going for it are either stupid or corrupt. I'm not actually copying your book, I'm training my printing press.
Silly billionaire, fair use is for people, not megacorps!
But seriously, fair use is for students, teachers, comedians, and all the people who can't pay or couldn't possibly pay enough for permission from copyright holders. A corporation training a closed model for profit can, and should, pay. If American copyright law is too onerous for megacorps to contend with, maybe it should be reformed.
Sorry, but no. If that was a statement was coming from an actual Open AI I would welcome it, but instead it comes from a pre-IPO "Open"AI, after it siphoned as much information as they possibly could.
This reminds me of a billboard I saw about 15 years ago when the state of Delaware was considering changing its laws to permit new casinos in addition to the three that had been grandfathered in when the restrictions went into place.
Delaware Park, one of those three casinos, put up a billboard opposing the bill, and it said, "No more casinos! THREE IS ENOUGH!"
It just goes to show that corporations will take any position -- even a casino taking an anti-casino position -- as long as it's self-serving.
From that article:
> The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
The New York Times lawsuit isn't purely about training on content - it's about memorization. They included examples of OpenAI's models being able to reproduce NYT articles entirely from the weights.
There's also stuff about RAG in there - about OpenAI (and Microsoft) features that run web searches and then summarize the results. I've not been following developments closely, I'd be interested to know how that part of the lawsuit is shaking out so far.
Right, it's not just that they saw the thing but also how they use the knowledge later.
To demonstrate this "what you do next matters" part, assume I get a book from the library and I read it, and I have perfect memory. So far, no problem, even with my fantastical memory. However what I do next matters:
A. I quote several short pieces of the book in a written critique. Fair use.
B. I use what I know to create a statistical breakdown of the author's favorite words and phrases. That's fine too.
C. I retype my own book which is 90% identical to the original and I publish it for money. Nuh-uh. Legal problems abound.
We can debate whether or not these companies are crossing the line, but the line does depend on what they do with the information, not just how they got it.
The reproduction argument was always the weaker of the arguments. The whole point of news is that it's new, if you ask ChatGPT for the current front page story that's not gonna be in the weights. They were able to contrive a single example of a very well known piece, but the number of people who will use that to circumvent the paywall is zero.
Lawfare Media typically has good content and analyses
https://www.lawfaremedia.org/search-results?prod_search-inde... (copyright, last 12 months)
https://www.lawfaremedia.org/search-results?prod_search-inde... (OAI, last 12 months)
I agree on this point!
And also, training an open weights AI on the outputs of an AI trained using "fair use", is itself "fair use" that benefits everyone.
Unfortunately congress and the senate will only pass malicious compliance laws like that when it's to the benefit of megadonors.
Honestly if you are putting models behind your own terms of use it's only fair use if I get to do the same for my content, but it's too late for me I folks in US sue them for every dollar or force them to make their models "fair-use".
I don't think scooping all of the internet and saying rules for thee but not for me is the way to go.
> rules for thee but not for me
Exactly, the same copyright rules should and do apply to both parties.
I'm glad copyright law didn't prevent the creation of LLMs. I'm also glad it doesn't prevent others from creating LLMs. I think that benefits us all.
Without government intervention, I think distillation will prevent these giant AI companies from recouping their investment. But that's not our problem.
That pretty obviously fails the copyright test. Training on LLM weights isn't transformative because you are producing...another LLM.
That pretty obviously fails the copyright test. Training on text/pictures/audio isn't transformative because you are producing...text/pictures/audio.
The other LLM is obviously transformative because it's a different LLM and therefore is not even the same kind of thing, like how a white horse is not a horse, obviously.
OK. Obviously the use-case has changed.
Consider a random example of fair use: print an excerpt of this three panel comic you would like to discuss.
Does building a business around the automation of fair use constitute fair use? I would say no.
The key point in all these AI training lawsuits is "transformative use".
If an LLM trains on NYT articles is the output of the model a substitute for the original article? Or will people continue to pay for the New York Times as before?
Whether they build a business around it doesn't really matter. Plenty of large businesses depend on fair use to avoid copyright lawsuits, including any tech company that hosts any kind of content (YouTube, Reddit, Meta). Even the media itself relies on fair use to be able to air any content.
> Plenty of large businesses depend on fair use to avoid copyright lawsuits, including any tech company that hosts any kind of content (YouTube, Reddit, Meta).
Those companies don't rely on fair use. They rely on the DMCA. There's an agreement with the user for use of content where the user controls the rights, and anything that users upload where someone else controls the rights, DMCA protects the site as long as they follow the procedures.
> Does building a business around the automation of fair use constitute fair use? I would say no.
This seems like an excellent and very clear distinction. Once it becomes a wholesale harvesting and processing for financial gain, it stops being fair use. Fair use clearly has a magnitude component.
Translation from corporate speak: when we steal from you that's good; when you steal from us that's bad
The irony of Sam Altman's company pretending to care about what's "fair" is so very rich.
He literally made billions exploiting the intellectual work of millions of his fellow humans, yet has the audacity to whine about how the system isn't fair.
There are a surprising number of people who think the only fair game is "heads I win, tails you lose".
Generational level of hypocrisy from OpenAI. What a nerve putting "authors" in the list first. Fair Use benefits OpenAI and the labs much more than anyone else.
Just be honest and don't try to blend it into "Everyone".
Fruitfulness of discussion might be aided by bearing in mind that US and European copyright law traditionally come from very different perspectives. Basically, societal utilitarianism vs work-is-an-inalienable-part-of-the-artist author's natural rights. "Copyright should last 5 years because starving keeps artists productive. And copyright owners have no say at all in their works being used for LLM training because it's good for society" is consistent with US foundations, at least pre-Berne. "Authors are not permitted to give up their right to personally refuse a use of their already-sold work, nor to give anticipatory permission for use in future technology" is IIUC unremarkable in a European context. An AI suggests[1] as philosophical background. People are coming at this from vastly different places. Bearing that in mind might help avoid talking past one another. IIUC.
I also note, when the internet was just catching on, and email was becoming less niche, but copyright law had not yet established sending email as an implicit grant of permission to make copies, most all use of email violated copyright. Network->disk->memory - all copies without permission. I'd get email with footers vaguely like "You are hereby granted permission to <do email> for the purposes of XYZ Corp - all other use constitutes infringement". There was a multi-year lag between accepted practice and legality. IIUC, forwarding email is still problematic today. Having lived through that, "Arg! <New tech> violates copyright!" gets less traction with me. And with copyright term extensions, and Berne on the US side, having a taint of unprincipled corruption, that also makes it less compelling.
[1] https://cyber.harvard.edu/people/tfisher/iptheory.pdf
Surely OpenAI would thus not be opposed to others distilling their models.
Is distillation fair use then?
I don't know why it shouldn't be as long as someone mixes in some additional post training.
It's against the TOS, so I would assume no
Why does OpenAI's TOS matter but everyone else's doesn't?
A private contract (terms of service) can't redefine federal copyright law, perhaps you're thinking of something else like "authorized access?" Ex:
1. I sneak into a movie theater without paying. Unauthorized access to the movie.
2. I write a review of the movie in the newspaper. Protected by fair-use, even if I sneaked into the theater.
Good question!
But let me guess, the trained model is a novel work and is protected, right?
Copying the exact weights? Yes.
But asking the trained model lots of questions, and creating a similar but different open weights model? No, that's fair use.
The current copyright rules cover both cases well and benefit everyone, I think. (I think copyrights are waaay too long, but that's another topic....) Whether trillion dollar companies can ever make a profit once open weights competitors distill them is, frankly, not our concern.
What about training on the source weights seeing as you can train on the source material under fair use?
> Authors, journalists, creators, academics, accessibility advocates, legal experts, librarians, economists and business leaders are voicing their support for the longstanding principle of fair use and its application to AI training.
Have you heard of the landmark $1.5 billion class-action copyright settlement in Bartz v. Anthropic?
Absolute cinema!
The Bartz v. Anthropic settlement wasn't about fair use, it was about pirating ebooks. The judge in that case said that the training on scanned, purchased books was fair use - Anthropic started doing that in 2024.
I have some notes about that here: https://simonwillison.net/2025/Jun/24/anthropic-training/
Would be the settlement bigger if the judge in that case said training on scanned albeit purchased wasn't fair use?
> Anthropic wins a major *fair use* victory for AI — but it’s still in trouble for stealing book
First line in your notes.
That was the title of the Verge article I linked to: https://www.theverge.com/news/692015/anthropic-wins-a-major-...
And it supports my summary here. Anthropic did "win" on the fair use issue, but lost on the pirating.
Note that I said "The Bartz v. Anthropic settlement wasn't about fair use" - and that's accurate, the settlement aka the $1.5bn was about the pirated books, even though the initial lawsuit was about fair use.
Here's my later post about that settlement: https://simonwillison.net/2025/Sep/6/anthropic-settlement/
ironic that they can talk out of both sides of their mouth, fair use is important for their training but if you distill it's definitely not fair use anymore
Are they just trying to rage bait everyone?
They must understand that this can literally only make people hate them more.
There is no open-ended "fair use" doctrine in the EU. Not everyone after all.
Bullshit.
Copyright originally maxed out at 28 years. Corporations rewrote the laws over and over, so that now copyright can last over 100 years, turning intellectual property into yet another rent extraction scheme.
Now, in this case, OpenAI argues that “fair use” is for the good of all. OpenAI doesn’t give a shit about what’s good for all. They want the ability to define intellectual property rules so that they can extract rent from someone else’s property.
If they want to benefit everyone, why aren’t their models open? For that matter, why is their name OpenAI, when they are not open? They’re publishing tons of work in mathematics, supposedly to benefit everyone. Instead, why don’t they forgo all research themselves and dedicate their work to supporting existing mathematicians in using AI effectively, so that the people that built the base of knowledge get credit for the new work?
Because OpenAI wants to extract the maximum shareholder value out of other people’s work.
Does anyone actually believe PR like this anymore?
I'm not sure what I find more morally repugnant to be honest. The largest wholesale theft of human labour ever seen, or the company that has made billions from this heinous act disingenuously and cynically trying to claim it as in any way "fair" (whilst simultaneously crying foul about distillation)
This is a wild re-imagining of what "Fair use" means in the US courts, which have a relatively straightforward four point test, and of which OpenAI, a commercial venture that produces exact copies of an article and thus acts as a market-replacement for the original work, fails handily.
What OpenAI wants is "free use".
LLMs trained on fair use are themselves fair use. Weights pls
What AI is doing is socializing the costs but privatizing the profits. This has been the model for every social media platform going back to forums. Everything after USENET, really.
If you're using the collective output of humanity to train your LLM then humanity should own the output. Problem solved.
One can easily understand this as the worker's relationship to the means of production.
What if I use the collective output of humanity to train my brain?
But you haven't. You've had maybe 40 years to read books, web pages, watch movies listen to music etc. Every LLM has millions of books, nearly every web page ever written, and every image available on the web. There's not even a rough order of magnitude comparison, your content consumption is not even a rounding error on the amount of text every LLM has in its weights.
Quantity has a quality all its own.
Isn't that worse, because each work contributes more to my brain?
I've had that same thought. It's a surprisingly valid point even though people tend to think you're joking. A human is capable of consuming a certain percentage of all human output, admittedly a low percentage, but that might change with, say, vastly extended lifespans in the future. What percentage is allowed?
The first difference is that somebody is typically getting paid for any content you consume. You buy books, TV shows, movies, magazines, whatever. Or you borrow them from a library who bought them. Or you pay a subscription for a streaming service. Or you use an ad-supported platform. If you bought any commercial IP and then reproduced it one million times so you can train a whole bunch of people at once, well that's a crime. So why isn't it a crime when AI companies do it?
The second is that you can only produce so much output. You can't be cloned a million times to produce a million times the output. Software can be.
LLMs are kind of like a billion people sharing one NYT account, 1 Netflix account and so on. And that's simply not sustainable.
I also believe that the lifespan of IP is way too long. Software shouldn't be patentable because it's ultimately math and math formulae aren't patentable. Other patents should generally have much shorter lives. Books, music, TV, movies should have much shorter copyright periods, say 10-30 years total.
AI companies ingesting all the IP on earth for free is like saying there is no IP at all. Some will agree with that. I don't. But again, I would have less of an issue with it if we as a society owned the output.
> The New York Times’ lawsuit, meanwhile, asks everyone to ignore a half century of legal history and progress, and toss aside the longstanding principle of fair use in their own narrow pursuit of an undeserved payday.
There's absolutely no "fair use" precedent that says that you can program a computer with copyrighted material. They're just asking for an exception essentially based on their compression being lossy, which of course the law takes the absolute opposite opinion on e.g. distributing a 1080p of my bluray disc has been determined to definitely not be fair use.
Their essential case is that they use the word "training," and people training themselves from books is fair use, so since they use the word "training" then what they're doing is fair use. It's stupid, and judges going for it are either stupid or corrupt. I'm not actually copying your book, I'm training my printing press.
Silly billionaire, fair use is for people, not megacorps!
But seriously, fair use is for students, teachers, comedians, and all the people who can't pay or couldn't possibly pay enough for permission from copyright holders. A corporation training a closed model for profit can, and should, pay. If American copyright law is too onerous for megacorps to contend with, maybe it should be reformed.
Sorry, but no. If that was a statement was coming from an actual Open AI I would welcome it, but instead it comes from a pre-IPO "Open"AI, after it siphoned as much information as they possibly could.
Whatever fair use is supposed to have been, generating never-before-seen amounts of revenue by hoovering up every word ever written is not it.
Not that much revenue yet but I feel you.