US Copyright Office found AI companies breach copyright. Its boss was fired

We are talking about the rights of the humans training the models and the humans using the models to create new things.

Copyright only comes into play on publication. It's only concerned about publication of the models and publication of works. The machine itself doesn't have agency to publish anything at this point.

these comparisons of llms with human artists copying are just ridiculous. it’s saying “well humans are allowed to break twigs and damage the planet in various ways, so why not allow building a fucking DEATH STAR”.

abstracting llms from their operators and owners and possible (and probable) ends and the territories they trample upon is nothing short of eye-popping to me. how utterly negligent and disrespectful of fellow people must one be at the heart to give any credence to such arguments

It's not only publication, otherwise people wouldn't be able to be successfully sued for downloading and consuming copyrighted content, it would only be the uploaders who get into trouble.

Do you have any links to cases where people were sued for downloading and consuming content without also uploading (eg, bittorent), hosting, sharing the copyrighted works, etc?

The problem isn't that people aren't aware that the scale and magnitude differences are large and significant.

It's that the space of intellectual property LAW does not handle the robust capabilities of LLMs. Legislators NEED to pass laws to reflect the new realities or else all prior case law relies on human analogies which fail in the obvious ways you alluded to.

If there was no law governing the use of death stars and mass murder, and the only legal analogy is to environmental damage, then the only crime the legal system can ascribe is mass environmental damage.

Does the distinction matter? If humans build a machine that uses so much oxygen that the oxygen levels on earth drop by half, can they say:

"Humans are allowed to breathe, so our machine is too, because it is operated by humans!"

Sounds like we’re talking about the right of AI company founders and people on HN to acquire wealth from creative works due to some weak argument concerning similarity to the human mind and creation of art. Since we’ve now veered into armchair philosophy territory, I think one could argue that the way human memory works and creates, both physically and mentally, from inspiration is vastly different from how AI works. So saying they’re the same and that’s it is both lazy and takes interesting questions off the table to squash debate.

It's a very consistently Silicon Valley mindset. Seems like almost every company that makes it big in tech, be it Facebook and Google monetizing our personal data, or Uber and Amazon trampling workers' rights, makes money by reducing people to objects that can be bought and sold, more than almost any other industry. No matter the company, all claimed prosocial intentions are just window dressing to convince us to be on board with our own commodification.

That's also why I'm really not worried about the "AI singularity" folks. The hype is IMO blatantly unsubstantiated by the actual capabilities, but gets pushed anyway only because it speaks to this deep-seated faith held across the industry. "AI" is the culmination of an innate belief that people should be replaceable, fungible, perfectly obedient objects, and such a psychosis blinds decision-makers to its actual limits. Only trouble is whether they have the political power to try to force it anyway.

Why do you think the obvious analogy is LLM=Human, and not LLM=JPEG or LLM=database?

I think you're overstating the legal uniqueness of LLMs. They're covered just fine by the existing legal precedents around copyrighted and derived works, just as building a death star would be covered by existing rules around outer space use and WMDs. Pretending they should be treated differently is IMO the entire lie told by the "AI" companies about copyright.

Yes, and then the response would be, "what have you done, we now need to pass laws about oxygen consumption where before we didn't".

Point being, laws aren't some God-ordained rules, beautiful in their fractal recursive abstraction, perfectly covering everything that will ever happen in the universe. No, laws are more or less crude hacks that deal with here and now. Intellectual property rights were questionable from the start and only got worse; they've been barely keeping up with digital media in the past couple decades, and they're entirely ill-equipped to deal with generative AI. This is a new situation, and laws need to be updated to cover it.

> The fatal flaw in your reasoning: machines aren't humans. You can't reason that a machine has rights from the fact a human has them. Otherwise it's murder to recycle a car.

The direction we're going, it seems more likely it'll be recycling to murder a human.

> these comparisons of llms with human artists copying are just ridiculous.

I've come to think of this as the "Performatively failing to recognize the difference between an organism and a machine" rhetorical device that people employ here and elsewhere.

The person making the argument is capable of distinguishing the two things, they just performatively choose not to do so.

LLMs are certainly not a jpeg or a database...

The google news snippets case is, in my non-lawyer opinion, the most obvious touch point. And in that case, it was decided that providing large numbers of snippets in search results was non-infringing, despite being a case of copying text from other people at-scale... And the reasons this was decided are worth reading and internalizing.

There is not an obvious right answer here. Copyright rules are, in fact, Calvinball, and we're deep in uncharted territory.

> LLMs are certainly not a jpeg or a database...

Their weights are derived from copyrighted works. Evaluating them preserves the semantic meaning and character of the source material. And the output directly competes against the copyrighted source materials.

The fact they're smudgy and non-deterministic doesn't change how they relate to the rights of authors and artists.

>>> Despite that, Humans can read a book, get inspiration, and write a new book and not be litigated against.

>> The fatal flaw in your reasoning: machines aren't humans. You can't reason that a machine has rights from the fact a human has them. Otherwise it's murder to recycle a car.

> We are talking about the rights of the humans training the models and the humans using the models to create new things.

Then that's even easier, because that prevents appeals to things humans do, like learning, from muddying the waters.

If "training the models" entails loading up copyrighted works into your system (e.g. encoded them during training), you've just copied them into a retrieval system and violated copyright based on established precedent. And people have prompted verbatim copyrighted text out of well-known LLMs, which makes it even clearer.

And then to defend LLM training you're left with BS akin to claiming an ASCII encoded copy of a book not a copyright violation, because the book is paper and ASCII is numbers.

> Yes, and then the response would be, "what have you done, we now need to pass laws about oxygen consumption where before we didn't".

Except in this case, we already have the equivalent of "laws about oxygen consumption": copyright.

> Intellectual property rights were questionable from the start and only got worse; they've been barely keeping up with digital media in the past couple decades, and they're entirely ill-equipped to deal with generative AI.

The laws are not "entirely ill-equipped to deal with generative AI," unless your interests lie in breaking them. All the hand-waving about the laws being "questionable" and "entirely ill-equipped" is just noise.

Under current law OpenAI, Google, etc. have no right to cheap training data, because someone made that data and may have the reasonable interest in getting paid for their efforts. Like all businesses, those companies would ideally like the law to be unfairly biased towards them: to protect them when they charge as much as they can, but not protect anyone else so they can pay as little as possible.

The problem is, you can say all of that for human learning-from-copyrighted-works, so that point isn't definitive.

>The fatal flaw in your reasoning: machines aren't humans. You can't reason that a machine has rights from the fact a human has them. Otherwise it's murder to recycle a car.

That might be true but I don't see how it's relevant. There's no provision in copyright law that gives a free pass to humans vs machines, or makes a distinction between them.

In the case of Copyright law, no provision means it will fall in "forbidden" land, not in "allowed" land.

Also in general, grey areas don't mean those things are legal.

Edit: this remains true even if you don't like it, ¯\_(ツ)_/¯.

>Copyright only comes into play on publication.

Nope.

You have a right to not publish any work that you own. This is protected by Copyright law.

Copyright covers you from the moment you create some sort of original work (in a tangible medium).

>In the case of Copyright law, no provision means it will fall in "forbidden" land, not in "allowed" land.

AI companies claim it falls under fair use. Pirates use the same excuse too. Just look at all the clips uploaded to youtube with a "it's fair use guys!" note in the description. The only difference between the two is that the former is novel enough that there's plausible arguments for both sides, and the latter has been so thoroughly litigated that you'd be laughed out of the courtroom for claiming that your torrenting falls under fai ruse.

> Under current law OpenAI, Google, etc. have no right to cheap training data, because someone made that data and may have the reasonable interest in getting paid for their efforts.

That's the thing though: intuitively, they do - training the model != generating from the model, and it's the output of a generation that violates copyright (and the user-supplied prompt is a crucial ingredient in getting the potentially copyrighted material to appear). And legally, that's AFAIK still an open question.

> Like all businesses, those companies would ideally like the law to be unfairly biased towards them: to protect them when they charge as much as they can, but not protect anyone else so they can pay as little as possible.

That's 100% true. I know that, I'm not denying that. But in this particular case, I find my own views align with their case. I'm not begrudging them for raking in heaps of money offering generative AI services, because they're legitimately offering value that's at least commensurate (IMHO it's much greater) to what they charge, and that value comes entirely from the work they're uniquely able to do, and any individual work that went into training data contributes approximately zero to it.

(GenAI doesn't rely on any individual work in training data; it relies on the breadth and amount being a notable fraction of humanity's total intellectual output. It so happens that almost all knowledge and culture is subject to copyright, so you couldn't really get to this without stepping on some legal landmines.)

(Also, much like AI companies would like the law to favor them, their opponents in this case would like the law to dictate they should be compensated for their works being used in training data, but compensated way beyond any value their works bring in, which in reality is, again, approximately zero.)

> Except in this case, we already have the equivalent of "laws about oxygen consumption": copyright.

Copyright laws were themselves created by the printing press making it easy to duplicate works, whereas previously if you half-remembered something that was just "inspiration".

But that only gave the impression of helping creative people: today, any new creative person has to compete with the entire reproducible cannon of all of humanity before them — can you write fantasy so well that new readers pick you up over Pratchett or Tolkien?

Now we have AI which are "inspired" (perhaps) by what they read, and half-remember it, in a way that seems similar to pre-printing-press humans sharing stories even if the mechanism is different.

How this is seen according to current law likely varies by jurisdiction; but the law as it is today matters less than what the law will be when the new ones are drafted to account for GenAI.

What that will look like, I am unsure. Could be that for training purposes, copyright becomes eternal… but it's also possible that copyright may cease to exist entirely — laws to protect the entire creative industry may seem good, but if AI displaces all humans from economic activity, will it continue to matter?

Agree. It feels a bit like earlier days in Bitcoin world. Eventually the courts decided how it was going to be and people like CZ had to pay a visit to jail, but there is now clear jurisdiction on that.

The same will happen with AI, no one will go to jail but perhaps it is ruled out that LLMs infringe copyright.

(Same thing happened in the early days of YouTube as well, the solution was stuff like MusicDNA, etc...)

> But that only gave the impression of helping creative people: today, any new creative person has to compete with the entire reproducible cannon of all of humanity before them — can you write fantasy so well that new readers pick you up over Pratchett or Tolkien?

That is even worse without copyright, as then every previous work would be free and you would have to compete with better works that are also free for people.

There were the famous napster cases, the kids and old ladies that got sued by the RIAA for using limewire to download some music.

There is also the fact that copyright holders will pressure your ISP into sending threatening letters and shutting off your Internet for piracy, even without you seeding. I haven't gotten the impression that you are in the clear for pirating as long as you don't distribute.

There's tonnes, this is a baffling question.

> That's the thing though: intuitively, they do - training the model != generating from the model, and it's the output of a generation that violates copyright (and the user-supplied prompt is a crucial ingredient in getting the potentially copyrighted material to appear). And legally, that's AFAIK still an open question.

It's still copyright infringement if I download a pirated movie and never watch it (writing the bytes to the disk == "training" the disk's "model", reading the bytes back == "generating" from the disk's "model").

> That's 100% true. I know that, I'm not denying that. But in this particular case, I find my own views align with their case.

IMHO, unless you're massively wealthy and/or running a bigcorp, people like you benefit a lot more from copyright than are harmed by it. In a world without copyright protection, some bigcorp will be able to use its size to extract the value from the works that are out there (i.e. Amazon and Netflix will stop paying royalties instantly, but they'll still have customers because they have the scale to distribute). Copyright just means the little guy who's actually creating has some claim to get some of the value directed back to them.

> and any individual work that went into training data contributes approximately zero to it.

Then cut all those works out of the training set. I don't think it's an excuse that the infringement has to happen on a massive scale to be of value to the generative AI company.

> Copyright laws were themselves created by the printing press making it easy to duplicate works, whereas previously if you half-remembered something that was just "inspiration".

Eh. I don't know the history, but my understanding was they were created because the printing press allowed others to deny the original creators the profits to their work, and direct those profits to others who had no hand in it.

After all, in market terms: a publisher that pays its authors can't compete with another that publisher that publishes the same works but without paying any authors. A word without copyright is one where some publisher still makes money, but it's a race to the bottom for authors.

> But that only gave the impression of helping creative people: today, any new creative person has to compete with the entire reproducible cannon of all of humanity before them — can you write fantasy so well that new readers pick you up over Pratchett or Tolkien?

Here's a hole in your thinking: if you like fantasy, would you be content to just re-read Tolkien over and over, forever? Don't you think that'd get boring no matter how good he was?

And empirically, "new creative [people]" manage to complete with Pratchett or Tolkien all the time, as new fantasy works are still being published and read. Do you remember that "Game of Thrones" was a mass cultural phenomenon not too long ago?

> That's also why I'm really not worried about the "AI singularity" folks. The hype is IMO blatantly unsubstantiated by the actual capabilities, but gets pushed anyway only because it speaks to this deep-seated faith held across the industry. "AI" is the culmination of an innate belief that people should be replaceable, fungible, perfectly obedient objects, and such a psychosis blinds decision-makers to its actual limits. Only trouble is whether they have the political power to try to force it anyway.

I'm worried because decision-makers genuinely don't seem to be bothered very much by actual capabilities, and are perfectly happy to trade massive reductions in quality for cost savings. In other worse, I don't think the limits of LLMS will actually constrain the decision-makers.

Nothing in copyright law talks about 'semantic meaning' or 'character of the source material'. Really, quite the opposite - the 'expression-idea dichotomy' says that you're copyrighting the expression of an idea, not the idea itself. https://en.wikipedia.org/wiki/Copyright_law_of_the_United_St...

(Leaving aside whether the weights of an LLM does actually encode the content of any random snippet of training text. Some stuff does get memorized, but how much and how exactly? That's not the point of the LLM, unlike the jpeg or database.)

And, again, look at the search snippets case - these were words produced by other people, directly transcribed, so open-and-shut from a certain point of view. But the decision went the other way.

>The fatal flaw in your reasoning: machines aren't humans.

I don't see how that affects the argument. The machines are being used by humans. Your argument then boils down to the idea that you can do something manually but it becomes illegal if you use a tool to do it efficiently.

>Their weights are derived from copyrighted works. Evaluating them preserves the semantic meaning and character of the source material.

That sounds like you're arguing that they should be legal. Copyright law protects specific expressions, not handwavy "smudgy and non-deterministic" things.

>The person making the argument is capable of distinguishing the two things, they just performatively choose not to do so.

I think that sort of assumption of insincerity is worse than what you're accusing them of. You might not like their argument, but it's not inherently incorrect for them to argue that because humans have the right to do something, humans have the right to use tools to do that something and humans have the right to group together and use those tools to do something at a large scale.

It's not about the tool, how you use it, or even how it works. It's about the end result.

I can go through and manually compress "Revenge of the Sith" and then post it online. Or, I can use a compression program like handbrake. Regardless, it is copyright infringement.

Can AI reproduce almost* the same things that exist in it's training data? Sometimes, so sometimes it's copyright infringement. Doesn't help that it's explicitly for-profit and seeks to obsolesce and siphon value from it's training material.

>that are also free for people

sounds like a good deal if you're people.

>Under current law OpenAI, Google, etc. have no right to cheap training data, because someone made that data and may have the reasonable interest in getting paid for their efforts.

If it were that cut and dried we wouldn't have this conversation at all, so clearly your position isn't objectively true.

The difference is we're humans, so we get special privileges. We made the laws.

If we're going to be giving some rights to LLMs for convenient for-profit ventures, I expect some in-depth analysis on whether that is or is not slavery. You can't just anthropomorphize a computer program when it makes you money but then conveniently ignore the hundreds of years of development of human rights. If that seems silly, then I think LLMs are probably not like humans and the comparisons to human learning aren't justified.

If it's like a human, that makes things very complicated.

>Sometimes, so sometimes it's copyright infringement.

So in those cases, the original authors might have a case. Generally you don't see these LLM doing that though.

>Doesn't help that it's explicitly for-profit and seeks to obsolesce and siphon value from it's training material.

Doesn't hurt either. That's a reason to be butthurt, but that's not a legal argument.

> That's a reason to be butthurt, but that's not a legal argument.

It is a legal argument, fair use specifically takes into account the intention. Just using it for commercial ventures makes the water hotter.

> A word without copyright is one where some publisher still makes money, but it's a race to the bottom for authors.

This is the case anyway; there are many writers competing for the opportunity to be published, so the publishers have a massive advantage, and it is the technology of printing (and cheap paper) that makes this a one-sided relationship — if every story teller had to be heard in person, with no recordings or reproductions possible, then story tellers would be found in every community, and they would be valued by their community.

> Here's a hole in your thinking: if you like fantasy, would you be content to just re-read Tolkien over and over, forever? Don't you think that'd get boring no matter how good he was?

The examples aren't meant to be exclusive, and Pratchett has a lot of books.

There's far more books on the market right now than a human can read in a lifetime. At some point, we may have already passed it, there will be far more good books on the market than a human can read in a lifetime, at which point it's not quality, it's fashion.

> And empirically, "new creative [people]" manage to complete with Pratchett or Tolkien all the time, as new fantasy works are still being published and read.

At some point, there will be more books at least as good as Pratchett, Tolkien, Le Guin, McCaffrey, Martin, Heinlein, Niven etc. in each genre, than anyone can read.

> Do you remember that "Game of Thrones" was a mass cultural phenomenon not too long ago?

Published: August 1, 1996 — concurrently with Pratchett.

Better example would have been The Expanse — worth noting that SciFi has a natural advantage over (high) fantasy or romance, as the nature of speculative science fiction means it keeps considering futures that are rendered as obsolete as the worn-down buttons on the calculator that Hari Seldon was rumoured to keep under his pillow.

If they were a database, they would be unquestionably legal, because they're only storing a tiny fraction of one percent of the data from any document, and even that data is not any particular replica of any part of the document, but highly summarized and transformed.

Llms can't express, that's the primary issue. You can't just make a collage of copyrighted works and shield yourself from copyright with "expression".

Scales of effect always come into play when enacting law. If you spend a day digging a whole on the beach, you're probably not going to incur much wrath. If you bring a crane to the beach, you'll be stopped because we know the hole that can be made will disrupt the natural order. A human can do the same thing eventually, but does it so slowly that it's not an issue to enforce 99.9% of the time.

Given that you can in fact prompt enough to reproduce a source image, I'm not convinced that is the actual truth of the matter.

It will when it inevitably hits their wallets. Be it via the public rejection of a lower quality product, or court orders. But both sentiments move slow, so we're in here for a while.

Even with NFTs it still was a full year+ of everyone trying to shill them out before the sentiment turned. Machine learning, meanwhile, is actually useful but is being shoved into every hole.

That's just the usual hand-wavy, vague "it's different" argument. If you want to justify treating the cases differently based on a fundamental difference, you need to be more specific. For example, they usually define an amount of rainwater you can collect that's short of disrupting major water flows.

So what is the equivalent of "digging too much" in a beach for AI? What fundamentally changes when you learn hyper-fast vs just read a bunch of horror novels to inform better horror novel-writing? What's unfair about AI compared to learning from published novels about how to properly pace your story?

These are the things you need to figure out before making a post equating AI learning with copyright infringement. "It's different" doesn't cut it.

Anyone writing "humans can learn from art why can't machines" or something to that effect is performatively conflating an organism and a machine.

My issue is with the rhetoric, if that isn't the rhetoric you are using I am not talking about you.

That's certainly an opinion.

My issue is that your rhetoric of "performatively conflating an organism and a machine" doesn't address the core issue of "humans can learn from art why can't machines". You're essentially saying that you don't like the question so you're refusing to answer it. There is nothing inherently wrong with training machines on existing data, if you want us to believe there is, you need to have some argument about what that would be the case.

Is your argument simply about your interpretation of copyright law and your mentality being that laws are good and breaking them is bad? Because that doesn't seem to be a very informed position to take.

>You can't just make a collage of copyrighted works and shield yourself from copyright with "expression".

And yet collage artists do that all the time.

>It is a legal argument

Not a very good one then.

My stated opinion is anyone who comes to an AI conversation and says "I can't tell the difference between organisms and computers" or some variation thereof does in fact have no trouble in practice distinguishing between between their child/ mom/ dad/ BFF and ChatGPT as is in fact questioning from a position of bad faith.

"There is nothing inherently wrong with training machines on existing data..." doesn't really conflate a machine with an organism and isn't what I'm talking about.

If you instead had written "I can read the Cat in the Hat to teach my kid to read why can't I use it to train an LLM?"

Then I do think you would be asking with a certain degree of bad faith, you are perfectly capable of distinguishing those two things, in practice, in your everyday life. You do not in fact see them as equivilent.

Your rhetorical choice to be unable to tell the difference would be performative.

You seem to think I'm arguing copyright policy. I really am discussing rhetoric.

I'll remind you that all fanart is technically in a gray area of copyright infringement. Legally speaking, companies can take down and charge infringement for anything using their IP thars not under fair use. Collages don't really pass that benchmark.

Yoinnking their up and mass producing slop sure is a line to cross, though.

I'm not an expert, but I thought fan art that people try to monetize in some form is explicitly illegal unless it's protected by parody, and any non commercial "violations" of copyright is totally legal. Disney can't stop me from drawing Mickey in the privacy of my own house, just monetizing/getting famous off of them.