Let me make a clarifying statement since people confuse (purposely or just out of ignorance) what violating copyright for AI training can refer to:
1. Training AI on freely available copyright - Ambiguous legality, not really tested in court. AI doesn't actually directly copy the material it trains on, so it's not easy to make this ruling.
2. Circumventing payment to obtain copyright material for training - Unambiguously illegal.
Meta is charged with doing the latter, but it seems the plaintiffs want to also tie in the former.
> Circumventing payment to obtain copyright material for training - Unambiguously illegal.
The judge in this case seems to disagree with you, not accepting the premise that downloading the material from pirate sites for this use inherently gets the plaintiffs an out from having to address fair use defense as to the actual use.
> the plaintiffs want to also tie in the former.
No, the defense wants to and the judge hasn't let the plaintiffs avoid it the way you argue they automatically can.
> The judge in this case seems to disagree with you, not accepting the premise that downloading the material from pirate sites for this use inherently gets the plaintiffs an out from having to address fair use defense as to the actual use.
This is a good point, as a reminder, the Folsom tests (failing or passing any one is not conclusive, they are to be holistically considered) are:
- the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes (Note also that whether or not the use is transformative is part of this test).
- the nature of the copyrighted work
- the amount and substantiality of the portion used in relation to the copyrighted work as a whole
- the effect of the use upon the potential market for or value of the copyrighted work
> If the former ever gets tested in court, it's the end of the road. All major AI companies have trained on copyrighted work, one way or another.
I can absolutely guarantee you that neither DeepSeek nor Alibaba's highly talented Qwen group will care even a little bit, in the long run. Not if there's value to be had in AI. (And I can tell you down to the dollar what LLMs can save in certain business use cases.)
If the US decides to unilaterally shut down LLMs, that just means that the rest of the world will route around us. Whether this is good or bad is another question.
The pattern hasn't changed in decades. Remember when ZTE copied Cisco's router code so precisely they included the same bugs and documentation typos?
LLMs are a drop on a hot stone compared to countless other factors why the world already is routing around the US - but I don't want to get political or economical.
These cases can set precedents that basically shut down all all future "useful" AI implementations if the judges go too far. You can bet that CCP doesn't care one whit about US copyright law if leads to them leapfrogging us, it will definitely be j-johna-jameson-laugh.gif . I think that's the point of the commenter from 20000ft
China found the perfect way to disrupt US tech, releasing open source versions of it for free or at least cheaper. Most of US tech is built on open source anyways and with the pace YC is investing in open source alternatives, it will win out in most niches.
My fear is that the US tech won’t be able to compete with state sponsored open source out of China and will move to ban open source or suppress it somehow.
Also, the Chinese work is legit. DeepSeek introduced a whole bag of new techniques like GRPO, and released quite a bit of good open source tooling.
And Alibaba's Qwen team seems to be quite genuinely talented at "small" models, 32B parameters and below. Once you get Qwen3 properly configured, it punches well above its "weight class." I'm still running real benchmarks, but subjectively, it feels like the 32B model performs somewhere between 4o-mini and 4o on "objectively measureable" tasks. It's a little "stodgy" and formal by default, though. We'll see what it looks like when people start fine-tuning it.
If the US dropped off the planet, it would maybe set LLM technology back a year.
Deepseek really changed how people think about Chinese tech. Even after new LLMs launched, Deepseek R1 and V3 hold their own on benchmarks and are significantly cheaper.
But seriously, LLMs are useless for anything except the most basic secretarial tasks and even there they still require as much reviewing as a high school intern. The reason executives want to use AI instead of humans is because AI is a capital expense that does not hit the financial the same way that wages do. (Capital expenses are excluded from EBITDA, which is the most common metric for measuring a company's financial performance. But it's become so popular for companies to push expenses to ITDA whenever they can that financial analysts are starting to push back and include ITDA in their analyses.)
In a nutshell, there are 3 ways to look at company financials: the PR way (EBITDA), the financial reporting way (GAAP, which includes ITDA), and the tax way (starts from GAAP or IFRS but with numerous rules on what items are included or excluded).
You point to Chinese companies disregarding any rules if there is value to be had in AI, while in the US, AI companies going to get 500 billion investment and a whistleblower is dead.
US AI companies will either make sure that a similar ruling will never be made or they will ignore it and pay the fines. They won't let anybody stop the gravy train.
Or AI companies could use some of their vast reserves of cash to pay for licensing agreements and pay people for their fucking intellectual property, then feed it to the beast.
But then they'd have to actually communicate with people and negotiate consent instead of just hoovering up everything they can get their hands on in their quest to replace it.
> If the former ever gets tested in court, it's the end of the road. All major AI companies have trained on copyrighted work, one way or another.
You assume that getting tested means the AI trainers lose, and also thar the model architectures that have been developed can’t be retrained from scratch with public domain, owned, and purpose-licensed material. (With several AI companies having been actively pursuing deals to license content for AI training for a while now.)
> If the former ever gets tested in court, it's the end of the road. All major AI companies have trained on copyrighted work, one way or another.
End of the road for major AI companies, and hopefully something better can be created once it's declared illegal without any murky waters.
There are LLMs trained on data that isn't illegally obtained, OLMo by Ai2 is one such model, that is actually open source and uses open data for training. Just because it's "very difficult" for OpenAI et al shouldn't be an argument to force them to behave ethically anyways. If they cannot survive acting legally, then so be it, sucks for them.
That would hardly be the end of the road. If copyright enforcement gets stricter then that will give a market advantage to the largest, best funded major AI companies like OpenAI because they can afford to simply buy licenses from copyright holders. I predict that we'll see new middlemen arise specifically to handle this licensing, much like the agencies that handle most music licensing today.
It's not the end, all these companies have "clean" datasets which they train their models on now, along with training on the previous "dirty" models. But it's been so many generations, that they don't need to worry about this copyright issue anymore
The lines for humans aren't clearly drawn, but they are drawn. The main difference is that humans are humans and LLMs are computer programs.
I see no reason why we should even entertain the idea of extending human rights to computer programs, and so far, nobody has been able to give me any good reasons why.
Furthermore, why are we only entertaining the human rights that can be used for profit-driven purposes? Why do LLMs, for example, not have the right to free speech? Or an attorney? It seems highly unethical to grant these computer programs some protections as if they're humans but not grant them personhood. This is akin to slavery, which is something we actually have to consider. Anthropomorphization is a double-edged sword. We cannot simultaneously consider them human when convenient and then consider them programs when it's not. Or, if we want to do that, we need to form coherent argument to why, how, and when.
You're thinking about it using the wrong framework IMO.
It's not about the program's rights, it's about the human's rights to use the program. Not the machine's right to do something, but the human's right to do something through a machine, or make a machine do something.
No, because the entire argument hinges on the fact that LLMs learn, which is like humans learning, so it's transformative. That only works if you consider learning or transformation to be something that does not rely on the human spirit. Which, actually, most people do not believe. And it's pretty difficult to argue - we don't even know how learning works for people.
A lot of people just jump to LLMs learning like it's a foregone conclusion. Mm... no. You need to convince people of that. You'll find if you talk to non-tech people, they're not just going to believe you when you say that.
Why isn't an LLM more akin to a database or a compression algorithm? Why is it closer to human learning? After all, humans are humans and we have the exclusive right and power to determine what is human and what isn't. And database and compression algorithms are computer programs, of the same kind as an LLM.
Then the results would be the same, and it would still be fair use. I have yet to see an example that demonstrates LLMs plagiarize by default or by tendency.
Your causality seems to be inverted here. You seem to be implying that "learning" (or the ingestion and retention of information for the same means) is banned by default for everything, but we decide to allow it for humans as the sole exception. This is not the case. Everything not prohibited is allowed, and "intermediate copies" are considered to be vital to fair use by the court system.
No, it wouldn't. Because if I record "Revenge of the Sith", compress it, and then distribute it for free online, that's obviously not fair use.
Fair use is pretty complicated. Part of Fair Use is the "The Effect of the Use on the Potential Market for or Value of the Work", which already puts even human commercial endeavors in a tough spot. You can make it work, but you have to really try. Satire like Weird Al or whatever isn't competing with the music it's satirizing, the venn diagram between those markets barely overlap. But a lot of LLM use cases are explicitly meant to obsolesce and siphon value from the things they used.
Like, why go to Getty Images when you could instead go to the glorified database, which has ingested all of Getty Images, and acquire an indistinguishable stock photo for free?
The only reason we're even really entertaining this is because people continually draw parallels to humans. You see, it's not stealing from Getty. It's more like if someone saw Getty Images and then went out and took a photo in that same flat, boring style. Except nobody saw anything. And nobody went out an took a photo.
> The only reason we're even really entertaining this is because people continually draw parallels to humans. You see, it's not stealing from Getty. It's more like if someone saw Getty Images and then went out and took a photo in that same flat, boring style. Except nobody saw anything. And nobody went out an took a photo.
But unless your argument is that the photo outputs from the GenAI are literally equivalent to the training data, you would agree the end result is the same, right? Anyone can see that the images are not the training data stitched together, so it doesn't even really matter how it all works mechanistically, even though your description ("glorified database") is wrong.
Part of my point is that you don't need to produce literally equivalent output. Again, if I record and compress "Revenge of the Sith", there's literally zero pixels shared between my recording and the actual movie. Cool, so I can go upload it for free then right? No, I can't.
Can GenAI produce indistinguishable images to what's on Getty Images? If you write the prompt correctly, yes. I know because there are services where you can get generated stock images.
> Part of my point is that you don't need to produce literally equivalent output. Again, if I record and compress "Revenge of the Sith", there's literally zero pixels shared between my recording and the actual movie. Cool, so I can go upload it for free then right? No, I can't.
That's because you would be redistributing the actual material, just in a really roundabout way. GenAI models are not that, they're not a database and don't work like one.
> Can GenAI produce indistinguishable images to what's on Getty Images?
That doesn't matter because you can't copyright a style. From the point of view of copyright law, it would look like you were copying nothing proprietary/owned at all.
> That's because you would be redistributing the actual material, just in a really roundabout way.
Right, which I’m arguing is what LLMs do just in an even more roundabout way.
The technical details of LLMs don’t actually matter. We don’t really care if they’re a database or not. The question is do they reproduce the source material? And yeah, pretty much they do, in a lot of instances. Not all, but a lot.
To produce yet another analogy, imagine I have a service X. You can pay and I will give you any movie you want. You don’t know how I do it. Is this copyright infringement or not? I would say yes. Now let’s say I reveal the secret - I open up photoshop and painstakingly recreate the movie frame by frame. I might make a mistake here or there. Is this still copyright infringement? I think it is.
> To produce yet another analogy, imagine I have a service X. You can pay and I will give you any movie you want. You don’t know how I do it. Is this copyright infringement or not? I would say yes. Now let’s say I reveal the secret - I open up photoshop and painstakingly recreate the movie frame by frame. I might make a mistake here or there. Is this still copyright infringement? I think it is.
Okay, but that is not what's happening here. Demonstratably so. The fact that a model is technically capable of overfitting to certain very repeated points in the training data doesn't mean the entire thing has to be shot down. The non-infringing uses far outweigh the offending ones, by a lot.
If what you say is true, and they do outright copy a lot, then it should be pretty easy for any IP holder to sue anyone who misuses the model that way for copyright infringement on those specific outputs.
> That only works if you consider learning or transformation to be something that does not rely on the human spirit.
Even changes made using simple non-ML algorithms can be transformative according to fair use doctrine, like the thumbnailing of images done by search engines. It's not meant in some spiritual sense.
The reason that’s okay is because you aren’t competing against the initial source material. A thumbnail on Google for “Revenge of the Sith” is not a replacement for watching the movie.
But, a lot of AI products are specifically and explicitly designed to obsolesce the thing they trained off. No need to go to Encyclopedia X or the NYT, this has the same content.
> The reason that’s okay is because you aren’t competing against the initial source material.
I don't mean to claim that search engine image thumbnailing is like-to-like in every consideration, just that it demonstrates there's no "human spirit" required in order to qualify as "transformative" as far as fair use is concerned. Search engine image thumbnailing has been found to be transformative, for instance in Perfect 10, Inc. v. Amazon.com, Inc.: "Google's use of thumbnails is highly transformative."
And, though I'm probably being pedantic here, I think it's important to distinguish that the other fair use factor you allude to is not whether you're "competing against" the original work, but specifically the effect of your use on the market/value of that original work. For example if your documentary uses a clip from a TV show and also happens to air in the same time-slot as that TV show - the extent you compete/displace market for the TV show in general (even as you would had you not included the clip of it) is not what's under consideration, but rather only the additional extent you displace its market specifically due to inclusion of that clip.
Because of that, I'd claim that some machine-learning-based tool that partially displaces the market for a work it was trained on (for instance, Google Translate displacing the market for a translated version of a book) might still be seen reasonably favorably under the market impact factor, so long as the extent it displaces that work is largely independent of whether it has trained on that work specifically (such as if the translation tool could already provide a decent translation of the original book even before having trained on its translated version).
The whole point of the fair use clauses is to protect humans. Clearly we can easily say that programs are altogether exempt in favor of humans, and it would be a proper thing to do, until the first real AI is built.
If corporations owned human slaves and fed them copyrighted materials so that they were inspired to produce original creative output, I don't think that creative output should enjoy legal protections either. Even if slavery were not illegal.
Because the obvious question would be - how can free people compete with that?
The FairTrained models claim to train with only public domain and legal works. Companies are also licensing works. This company has a lawful, foundation model:
So, it's really the majority of companies breaking the law who will be affected. Companies using permissible and licensed works will be fine. The other companies would finally have to buy large collections of content, too. Their billions will have go to something other than GPU's.
Not really sure a claim is good enough. I don't know that you can just go into court and say, "Trust me, I don't use copyrighted material."
And I also can't see any way, other than providing training data and training an identically structured model on that data, that a company can conclusively show that they got the weights in an allegedly copyright free model from the copyright free training data a company provides.
While the others are correct, I'm with you in the sense that I don't know if what they claim is true. I've also found others, like one in Singapore, that didn't use it on data that was as legal as news reports claimed. It might turn out to have problems.
There is benefit to using them, though. For one, they've tried really hard to be legal. That sets a positive example, shows good faith if they were sued, and reduces risk for those using them (good faith on our part). Also, one can be sure that they can ditch or replace any outputs in the long term if they're ruled illegal. So, we try not to use the A.I.'s in a way where losing access to them seriously damages our business.
That's the best I can offer until legal reforms happen.
If training, one can train it in Singapore on material you he or she has legal access to. Their law pretty much let's you use anything for AI purposes so long as you legally can access it yourself. To further reduce the risk, they should crawl it themselves, too, taking care to avoid risky sources.
Civil courts work by you proving damages (at least in the USA), not by you going on fishing expeditions because they "might" have done something.
So good luck finding the thing that looks exactly like your copyrighted work that's not in the corpus, if you can yeah, you might be able to prove it.
At the end of the day its like a lot of business, where a liability shell game is played out, and if the chain of evidence cant be drawn quite brightly then lawsuits would be frivolous at best.
“Copy” is ambiguous here. Of course data is copied during training. That said, OP is referring to whether the resulting model is able to produce verbatim copies of the data.
Why does it have to be verbatim? Seriously, this I don't understand.
If I produce a terrible shakycam recording of a film while sitting in a movie theater, it's not a verbatim copy, nor is it even necessarily representative of the original work -- muddied audio, audience sounds, cropped screen, backs of heads -- and yet it would be considered copyright infringement?
How many times does one need to compress the JPEG before it's fair use? I'm legitimately curious what the test is here.
The purpose of copyright is it progress the arts and sciences. Not to guarantee profit. Guaranteeing profit is just the way we encourage people to progress the arts and sciences.
That is why so called derivative works are allowed (and even encouraged). If copyrighted material is ingested, modified or enhanced to add value, and then regurgitated that is legal, whereas copying it without adding value is not legal.
If derivative works weren't deemed acceptable copyright would have the opposite of it's intended effect and become an impediment to progress.
> That is why so called derivative works are allowed (and even encouraged).
Derivative works are not given a free pass from the normal constraints of copyright. You cannot legally publish books in the universe of A Song of Ice and Fire without permission from the author (and often publisher), calling them “derivative works.”
It’s why fan fiction is such a gray area for copyright and why some publishers have historically squashed it hard.
The exceptions for this are typically fair use, which requires multi-factor analysis by the judiciary and is typically decided on a case-by-case basis.
> That is why so called derivative works are allowed (and even encouraged).
Derivative works are not "allowed (and even encouraged)" without a license from the copyright holder. Creating a derivative work is an exclusive right of the copyright holder just like making verbatim copies and requires a license for anyone else, unless an exception to copyright protection (like fair use) applies.
Derivative works are not generally allowed in many jurisdictions. Try releasing a cover song without clearing it first etc. Even using a recognizable sample will bite you
Derivative works are tolerated in some cases like some manga or fanfics but it is a gray area and whenever the author or publisher wants to pursue it it is their full right to do it. Many do pursue it
(You can get inspired by something, and this is where some arguments can happen if you get inspired too mmmm literally, but no one will say with a straight face that inspiration is a thing that happens to software)
> Try releasing a cover song without clearing it first etc. Even using a recognizable sample will bite you
So… it’s complicated. This is one of the weird areas where music copyright and other copyright seem to differ in the US.
In the US the situation is complex and there are a lot of weird special interests [0], but generally a composer/author of a song has the right to decide who first records and releases the song, but after the first recording covers require a mechanical license, which is compulsory (ie: the author cannot object).
In music there are _a lot_ of special cases and different rights are decided with different kinds of licenses, some of which are compulsory. I think it’s an area that doesn’t make for good analogies with copyright in other media.
> covers require a mechanical license, which is compulsory
Which is compulsory for the performer too.
A derivative work like cover is sort of acceptable when it's performed by a person live for some audience (grey area but twitch sort of allows it. with a bunch of rules). As soon as you want to publish it you MUST have a license. And chatbot is a derivative work totally not performed live by a person for some audience
I saw great tracks that got taken down from all legal channels because they featured a sample from another song. Sometimes they remained up but mostly they were taken down. It is fully original publisher's discretion...
You are absolutely right. I should have phrased that differently. Derivative work is a legal term, but I misused it above. I should have either used another term or been clearer.
If the work is "derivative" in the legal sense it is copyrighted, and you may not create derivative works without the copyright holders permission.
What I should have said is that simply being inspired by a work or copying unprotectable elements (like facts or ideas) does not create a derivative work.
For example, if ChatGPT were to generate Star Wars, except with Dookies instead of Wookies, that might be illegal. If it were to learn what a spaceship is from Star Wars and then create something substantially new it would not. The key is is that it must not be substantially similar to the original. You must add enough value that it becomes something new, not just rehash the original.
Learning is like inspiration, it's something humans do. Don't get fooled by "learning" in "machine learning", it doesn't mean machine becomes like human in legal sense...
There’s something called a substantive transformation test in copyright law. When you write a summary of a book, you don’t infringe on copyright because it’s a “substantial transformation”. This goes along with the idea that you can copyright the text but not the ideas it expresses.
When model training reads the text and creates weights internally, is that a substantial transformation? I think there’s a pretty strong argument that it is.
The point here is that book files have to be copied before they can be used for training. Copyright texts typically say something like "No unauthorised copying or transmission in any form (physical, electronic, etc.)"
Individuals who torrented music and video files have been bankrupted for doing exactly this.
The same laws should apply when a corporation downloads torrent files. What happens to them after they're downloaded is irrelevant to the argument.
If this is enforced (still to be seen...) it would be financially catastrophic for Meta, because there are set damages for works that have been registered for copyright protection - which most trad-pubbed books, and many self-pubbed books, are.
Only if they seeded the data and some other entity downloaded it, i.e. they hosted the data. In a previous article I believe it was called out that Meta was being a leecher (not seeding back what they downloaded).
It's the hosting that gets you, not the act of downloading it.
I would like to expand on this, since it seems to be a common misunderstanding. Lets imagine a hypothetical situation where one friend loans a book to another, who then makes a copy of it.
The lender owns the book, and it is within his rights to loan it to whoever he wants. That is legal. Making this illegal would end libraries.
The borrower is well within his rights to accept the book, and as the current owner he is even allowed to make a copy of the book (see the famous TIVO case). Making this illegal would end backups and format/time shifting.
When the borrower returns the book, he keeps the copy. Oh no! Surely he must now become a criminal? Nope. Possessing an unauthorized copy is also not illegal, despite what many copyright holders would like you to believe. Making this illegal would also criminalize a lot of legitimate format/time shifting, again see the famous TIVO case.
If the borrower were to loan his homemade copy to someone else THEN it would finally become illegal.
Don't read too much into what I am saying. I am not even talking about the AI piece.
I download a torrent with movie that I didn't pay for. If I don't allow to seed it, then I don't get in trouble. If I let it seed either during the download process or after, I'd get a DMCA notice if that torrent/magnet link was getting tracked.
I don't need a hypothetical book, that is just how it works if I were to download illegally obtained documents/media.
As technical as people are in this thread, easy to tell when folks didn't have their parents wondering why they were getting scary letters from the ISP.
If you made a durable copy of a book in your example to keep for yourself and use later that's already a grey area. But no one does it with books. People do it with other media tho, and big surprise get prosecuted for it. As you may know, in developed countries people get served notices for torrenting
But if you make books contents available online via some service that regurgitates its contents you would be totally sus because you can be considered in a business of selling derivative works.
There was a relatively famous Google case regarding their digitization of books without the authors consent in 2015. Although it's not a perfectly analogous to this situation.
In Googles case they were digitizing the books (that they did not own), and publishing snippets for search users to help them find books and other material that weren't indexed on the web. The court found they had that right, but did place some pretty strict limits on them.
Still, Google was allowed to keep their database of scanned material despite not owning the originals.
According to your link and this comment https://news.ycombinator.com/item?id=43899406 Google's scanning project was ruled fair use because it was "transformative" and didn't harm the market for the works. It allowed searching within books that was otherwise impossible at the time, but by not providing the full text of the books it didn't meaningfully reduce sales.
Someone photocopying a book to read on the toilet (and leave the original on their nightstand) isn't engaging in transformative use. They're also harming the market for the work because if they hadn't made this photocopy, they would've had to buy a second copy of the book to get the same benefit.
> Someone photocopying a book to read on the toilet (and leave the original on their nightstand) isn't engaging in transformative use.
The situation you describe is more akin to "format shifting" or maybe "space shifting", which is converting copyrighted material to other formats or places as a backup, for preservation purposes, or just for convenience. It is legally protected in the US, and most of the rest of the world. I believe that was settled due to early litigation with VCR's, but there are mountains of relevant cases at this point.
I do recall reading that the UK recently changed their laws in that regard, and like most copyright law it can depend on the specifics of the situation and even (*yuck*) intent. So it's always worth checking if you're unsure.
This is a leap in the argument. We've gone from the right to use a work to "unless the result is identical or close to it, we have full rights to all works.".
Even if you argue the LLM's are merely summarizing content, they still had to illegally download that content in the first place. The model can't read and simmarize the texts unless the text was illegally downloaded and copied. Piracy isn't suddenly legal just because you promise to delete the movie you downloaded after watching it.
That's not really true. Models train (in greatly simplified way) by being shown an excerpt and being told to guess the next token from the excerpt. They push around their weights until the token they output matches the next token in the excerpt. Then the excerpt is no longer needed. You can think of it like the article is loaded, the LLM plays this token guessing game through it, then the article is discarded. On the face of it this is what happens, but it gets hairier depending on how exactly this process is done. But it is seemingly not far removed from how humans consume content (acquire, read, discard), hence the legal blur.
The is a bar of usage built into the law, otherwise everyone who reads this wired article is violating copyright by making a full copy to their computer. Generally making non-lasting copies is fine, otherwise the internet wouldn't work.
AI doesn't need lasting copies to train, however I don't know what the actual implementation is. But if it's ruled that they can only use copyrighted data if it's not stored for more than the time it would take a human to consume, It wouldn't really cripple the models, but perhaps make training more logistically challenging.
It's important to understand that models are not data archives. They are statistical constructs made from getting quizzed, that uses human made content to generate the quiz questions.
> Except that humans don't make full copies to RAM, or disk or paper.
Images on your retina form exact copies.
They are scanned and translated into impulses that are then sent to a first set of "neural columns" - that's an exact copy.
This is then connected to the visual cortex by the two most high bandwidth links in the human body ("the optical nerve", there's 2 of them of course, always wondered why everybody insists on using the singular). Why would you have that high bandwidth link unless to create verbatim copies.
The way those columns are structured also very strongly suggests they make carbon copies, which they then make available on the "brain bridge" (which is probably at least vaguely similar to the "attention matrix" of a transformer). If it does work like that, that's also a verbatim copy.
The only way "humans don't make full copies to RAM" is that humans don't have separate RAM. The processing power is colocated with the processing, even on a microscopic level. You know, what everybody knows is the best way of doing things even in silicon, it's just incredibly impractical if you can't rebuild your circuit every time there's a slight change to the instructions your "computer" carries out (the brain is not a "Von Neumann architecture", except it kind of is when it regrows connections. But in the short term it isn't)
Sure they can. You can learn a copyrighted work by hard, even indirectly, then quickly duplicate it by hand. Mozart was originally famous for making a business out of that.
It is different thing. When you copy data into computer's RAM, that might be copying as defined in law [1]:
> Using software almost always involves creating copies, even though many of these copies only exist for a very short time. For example, executing a program means copying it from the hard disk into RAM so that the CPU can interpret the instructions. Because of this, the right to run a program is considered to fall under the copyright of the author.
For comparison, when a human looks at the letters, there is no copying.
Also, models can reproduce text verbatim which proves that they store it.
So it is unfair when ordinary folks got sued for this and Zuckerberg wants to get away with a million times larger violation. He must go directly to jail.
completely different scenarios. A pirated movie is marketed/sold as a copy of something, which is not fair use. An LLM just remembers/get inspired by what it consumes
I don't believe that's correct. The existence of filters to block potentially copyrighted materials contained in LLM outputs proves that they don't just "get inspired."
It seems like it is very much a matter of fidelity.
> An LLM just remembers/get inspired by what it consumes
As mentioned in another comment, LLMs (and most popular machine learning algorithms) can be viewed, correctly, as compression algorithms which leverage lossy encoding + interpolation to force a kind of generalization.
Your argument is that a video wouldn't count as pirated if the compression used for the pirated copy was lossy (or at least sufficiently lossy). The closest real world example would be the cases where someone records a the filming of a movie on their phone then uploads it. Such a copy is lossy enough that you can't produce anything really like the original, but my most definitions is still considered copyright.
They are in no way compression algorithms. They can be modeled like that in the same way you can model humans as lossy compression algorithms.
You would never use a human to backup your financial reports, but the human might be able to give a good overview. You would never use an LLM to backup your financial reports, but they might be able to give a good overview.
AI training data is disposable. There is nothing that could be called a compression algorithm that disposes all of the data you put into it. AI uses training data as examples of what the next token in a token sequence is. The examples are disposable reference points, not the model itself. That's how you get image models that are 20GB in size despite training on 20PB of data. It's 20PB of examples used to form the shape of a 20GB model. You could show it 5GB of training data or 500EB of training data and it would still be 20GB - because it is not a compression algo, it's a 20GB shape formed by external data.
>It's 20PB of examples used to form the shape of a 20GB model. You could show it 5GB of training data or 500EB of training data and it would still be 20GB - because it is not a compression algo, it's a 20GB shape formed by external data.
You can compress 20PB of text to 20Gb or even less, if input is super repetitive.
So the same with images, if 50% of the images are cats then you learn how to represent the cat pixels with a few vectors and then you could represent all the cats int he world doing all possible cat actions.
But please have the courage to respond to this,
when the AI is caught regurgitating the exact text from a popular book, the exact verses from a poem, the exact code function from some code , then how can you defend that is not memorizing things? If a human uses my poem(after they read it) and signs his name under it would you defend them?
The point is that it isn't compression. Its molding a plain structure iteratively into a ultra complex one. The model starts and ends at 20GB. It might have features that are reminiscent of compression or act like it, but under the hood there is nothing like zip, rar, H.265, or JPEG going on.
And yes LLMs can recall exact material, but it is excerpts and fragments. There is statistical significance to it's ordering. Humans readily do this too (excerpts and fragments), most artists can draw a batman symbol (but not an episode of batman). That doesn't in anyway mean that artists should not be allowed to ever see a batman symbol. It means that artists shouldn't be allowed to get paid to draw one. And they are not. And LLMs are not exempt either.
But the fix is output filtering, just like everything else that can violate copyright. Which is already being done (albeit poorly, but way better than 2 years ago), the same as artists will not draw the batman symbol for you despite being able to.
OK, so we name it something different, you transform inputs into smaller outputs.
If I make a script without AI that transforms someones poems without permission so sometimes it outputs the exact poems but sometimes it does it wrong, when is my script fair to use and when what I did is illegal. say my script contains words and matrixes of numbers the original poems are not directly inside, the script transformed them into vectors.
maybe even simpler, I create a zip format where I randomly replace words with their synonym, or group of words with something equivalent. Would you defend this as original ? why my random transformations are not original while mathematical transformations you will defend ?
And how can you suggest putting output filters to protect only the giants for copyright and everyone else gets screwed.
YouTube actively filters small channel copied content too.
We should be building robust copyright filters and everyone should be able to contribute their work to it.
but that is a different issue than whether or not an LLM is legally allowed to view a work that is publicly available.
Again, pretty much every artist is capable of off-hand copyright violation on the spot. This has been true forever. We don't bar them from seeing art to prevent this.
>but that is a different issue than whether or not an LLM is legally allowed to view a work that is publicly available.
We are not talking about the scenario where LLM does a web search and "transforms" your blog as a response to a user prompt.
We are talking about people writing code to torrent content, including copyrighted content as books and newspaper articles then this same people create soem software/script/blackbox and they put this copyrighted stuff inside and this same people also put a filter in the output to attempt and block SOME of the copyrighted material to be spitted out(they did it for Dune, Harry Poter , probably other popular stuff but for sure they did not done it for Romanian copyrighted material or some small writer).
If I publish a poem I do not give implicit permission to anyone to use it as they see fit, and have scripts/software/backboxes train on it and spit parts of my poem out.
I'm sorry, but this a fundamentally incorrect view of machine learning (including, but not limited to transformers).
From an information theoretic perspective the two are essentially identical with the exception that standard compression algorithms do not have a proper "loss" function other than just trying to minimize reconstruction loss with the resulting compression size.
Here's a link to the section on the Wikipedia for more information if you'd like [0]. MacKay's Information Theory, Inference and Learning Algorithms is the standard full text treatment of this topic [1]. Ted Chiang's article "ChatGPT is a Blurry JPEG of the web" is pretty good "pop sci" exploration of this topic if you don't want to get too into the mathematics [2].
>They can be modeled like that in the same way you can model humans as lossy compression algorithms
Humans are totally capable of data compression. This will just devolved into a semantics game of what a data compressor is.
LLMs were not developed to be, do not function as, and are not use as data compression utilities. Please, come knocking when a service provider exists that will use LLM's to compactly store your company data.
> It has long been established that predictive models can be transformed into lossless
compressors and vice versa. Incidentally, in recent years, the machine learning
community has focused on training increasingly large and powerful self-supervised
(language) models. Since these large language models exhibit impressive predictive
capabilities, they are well-positioned to be strong compressors. In this work, we
advocate for viewing the prediction problem through the lens of compression and
evaluate the compression capabilities of large (foundation) models. We show
that large language models are powerful general-purpose predictors and that the
compression viewpoint provides novel insights into scaling laws, tokenization, and in-context learning. For example, Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively.
> LLMs were not developed to be, do not function as, and are not use as data compression utilities.
Again, from a information theoretic view point, this is exactly what they are doing, how they where developed and how they function.
I don't know any serious researcher in ML that would find this claim even remotely controversial. It's really not just "a semantics game", its a part of a foundational understanding of the topic. If you want to understand LLMs from this perspective, a good place to start is with an auto-encoder which does try to learn a standard compression algorithm, the move on to more sophisticated embedding models (found in a lot of recommender systems) which try to learn an additional objective on top of minimizing reconstruction error. You'll then see that Transformers and all other major NN architectures fall out of these basic principles.
> Please, come knocking when a service provider exists that will use LLM's to compactly store your company data.
This is literally what every vectordb company does right now, as well as all "chat with your docs" type startups.
VectorDBs are not LLMs or SQL replacements and RAG is not data compression. Again this is just going to dwindle into semantics and where one draws boundaries. I can randomly remove bits from my HDD and call it compression. If you think humans are data compressors then I have no argument.
Can you get transformers to regurgitate information verbatim? Yes.
Would anyone in their right mind rely on a transformer to do so? No.
Would anyone in their right mind rely on a vectorDB to do so? No.
Would anyone in their right mind use a vectorDB/RAG/SQL/transformer combo to do so? Yes.
Is youtube going to drop VP9 for GeminiEncode to save google billions in bandwidth? No.
The test is if a judge says it is fair use, nothing else.
The judge will take into account the human factor in this matter, e.g. things like who did the actual work, and who just used an algorithm (which is not the hard part anymore, code can be obtained on the internet for free). And we all know that DL is nowhere without huge amounts data.
Half the point of this crap is labor automation. How does it not compete? If I write a bunch of books and then you use my work to make a machine that writes books in my style, you are using my labor to directly compete with me.
> If you read a book and later understand its plot but can only explain it in your own words, did you copy it?
I think that is the center of the conversation.
What does it mean for a computer to "understand"?
If I wrote some code that somehow transformed the text of the book
and never return the verbatim text but somehow modified the output,
I would likely not be spared
because the ruling will likely be my transformation is "trivial".
Personally, I think we have several fixes we need to make:
1. Abolish the CFAA.
2. Limit copyright to a maximum of 5 years from date of production with no extension possible for any reason.
3. Allow explicit carveout in copyright for transformational work. Explicitly allow format shifting, time shifting, yada yada.
4. Prohibit authors and publishers from including the now obviously false statements like "No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording" bla bla bla in their works.
5. I am sure I am missing some stuff here.
For brand protection, we already have trademark law. Most readers here already know this but We really should severe the artificial ties we have created between patents, trademarks, and copyright.
I just happen do read the Phoenix Technologies wikipedia page a few days ago. This company is known for developing BIOS software for computers. Maybe you've seen their logo when you first turn on your computer.
In early computing, everything was closed sourced. Quoting the wikiepdia page,
To develop a legal BIOS, Phoenix used a clean room design. Engineers read the BIOS source listings in the IBM PC Technical Reference Manual. They wrote technical specifications for the BIOS APIs for a single, separate engineer—one with experience programming the Texas Instruments TMS9900, not the Intel 8088 or 8086—who had not been exposed to IBM BIOS source code.
The legal team at Phoenix deemed inappropriate to "recall source in their own words" for legal reasons.
My non-legal intuition is that these companies training their models are violating copyright. But, the stakes are too high--it's too big to fail if you will. If we don't do it, then our competitors will destroy us. How do you reconcile that?
You're right that if we want to have usable LLMs at all, there's no way around training them on copyrighted materials. So it has to be allowed, but in a way that compensates the original authors somehow. For example, every model provider has to publicly declare all works used for training, and then all inference providers offering that model have to collect a per-token tax that gets distributed to authors in proportion to their presence in the dataset (by the by, this could also be a way to fund websites like Wikipedia).
But any such arrangement needs to be hammered out by the legislature. As laws are, I think it's pretty clear that infringement is happening.
Perhaps Phoenix just looked at the potential adversary (IBM) and decided to approach the project in an exceedingly cautious way, knowing that IBM could litigate it forever if there were any plausible argument that they "copied" even a line of code.
To what extent connections in a neural network are analogous to connections between neurons in your brain is open to interpretation and study, but the point of the analogy is that in neither case is a copy being made.
I can arrange a series of bricks in many ways to try and build a wall but that doesn't mean I will automatically get a good result if my process (like a ML training algorithm) doesn't precisely arrange then in a manner that produces a rigid wall with the desired characteristics. In the same vein you can have a fancy neural network arranged by some fancy LLM training algorithm with gobs of data about a subject but current methods likely won't produce anything with the depth of "understanding" that a human can do. It's a crumbly wall that falls once you do any real inspection or put any real load into it.
Yeah but a copy IS made. A human just reads. The machine copies the full text then compresses a lossy copy in its weights. You keep dodging that with tortuous analogies of a human learning.
I’m sure all these ‘clever’ questions would be useful if this trial was about humans but it’s not.
Model training works roughly by feeding the model a text excerpt and then hiding the last word in the excerpt. The model is then asked to "guess" what the final word is. It will then move around it's weights until the guess sufficiently matches the actual token. Then the process repeats.
The training material is used to play this guessing game to dial in it's weights. The training data is picked up, used as reference material for the game, and then discarded. It's hard to place this far from what humans do when reading, because both are using the information to mold their respective "brains" and both are doing an acquire, analyze, discard process.
At no point is training data actually copied into the model itself, it's just run past the "eyes" of the model to play the training game.
The obvious test for this would be to have the models produce an article from before and after its cutoff date and see if the output is still verbatim.
It would be a remarkable quirk of statistics that, if given all text on the internet except for the NYTimes back catalogue, a model would produce any NYT article.
"Of course data is copied during training" is copying. As far as I know, the law is consistent that temporary copies are also covered by the copyright act, and that's how some analogous cases were resolved.
FWIW the essence of copyright law AIUI is: copying is not permitted, unless done in a form explicitly allowed by the license holder.
This scenario seems quite contrived but is there an actual court precedent allowing it? I'm 100% confident no one will ever prosecute you for doing it but that's not the same thing as "allowed".
This case was about pointing <recording equipment> at <content someone is allowed to access> and making a copy to transmit it to <only that person>. The Supreme Court held it to be illegal. There was a lot of money on the line, which is why it went so far.
> there is also a principle that you can have ephemeral copies as part of handling and processing the work.
I'm not aware of this principle. Where is it spelled out?
> Also the result of that case was self-contradicting
I agree the verdict was a travesty. An innovative business went to ridiculous lengths to stay on the right side of the copyright mafia (data centers with tiny individual TV antennas for each subscriber FFS!) while providing a better product and experience. They still had the hammer brought down on them.
> I'm not aware of this principle. Where is it spelled out?
Well, do CDs give you a license agreement that allows you to copy the data? I've never seen one. And it's nearly impossible to play a CD without that copying.
> And it's nearly impossible to play a CD without that copying.
Exactly. It's how you are supposed to use the CD. That's not true for your example of a book on a webcam. You're supposed to read the book, not an image of the book.
That's not the same thing as copying for "processing".
This is not correct. It's only true that no one will go to the effort of prosecuting you for keeping photocopies of books in your home. But copyright law doesn't allow you to do it.
> That said, OP is referring to whether the resulting model is able to produce verbatim copies of the data.
While a tool being used to create infringing copies of some other work (whether or not it is the source material used to create the tool, and whether or not the infringing material is also verbatim copies) is relevant to whether the tool vendor is liable for contributory infringement for the infringing use of the tool, the absence of a capacity for creating such copies isn't usually enough to say that copying to make the tool isn't infringing.
(That said, generative AI tools, including LLMs specifically, have been shown to have the capacity to make such copies, to the extent that vendors of hosted models are now putting additional checks on output to try to mitigate the frequency with which verbatim copies of substantial portions of training-set works are produced, so arguing that LLMs can't do that is silly.)
> LLMs specifically, have been shown to have the capacity to make such copies
Exactly. I asked my Gemma how long of a quote it could give me of a given book, if I were the author & gave express permission, and I was a bit surprised it readily admitted it could
> Without Permission (Current Limit): Single sentence.
> With Broad Permission (Full Reproduction Allowed): I could theoretically quote the entire book.
> That said, OP is referring to whether the resulting model is able to produce verbatim copies of the data.
Transformers are fundamentally large compression algorithms where the target of compression is not just to minimize reconstruction loss + compressed file size. In fact, basically all of machine learning used today can be viewed through the lens of learning a compression algorithm with added goals other than the usual.
By this logic if I create a lossy Jpeg of a copyrighted image it's not "copying" because the lossy compression.
So if they could produce verbatim segments, that would be a violation? The technology is certainly there and these companies need to work backwards to prevent that.
The US Federal government operates with the rule that if human eyes don't look at it it doesn't count as a copy or looking at it. This allows them to unconstitutionally spy and log all people's telecommunications. Applying it here it seems pretty clear that corps are within the established bounds. As are any human persons that want to train an LLM this way.
Someone engaged in large-scale unconstitutional spying does not give two fs about incidentally doing some copyright violations to achieve the spying. These are entirely orthogonal considerations.
A good way of thinking about this is: consider the case where the data in question is illegal. Could you get into trouble for not only having access to it but also making copies of it?
I would argue that as an individual, real person, obtaining content without a license and personally consuming that content is significantly different than a corporation doing the same. My rational is that distribution of that content is (or should be) the primary offense. If I work for a company and they direct me to collect a bunch of content without a license and then I pass that to other members in my team to train a model, I've now distributed that content at the direction of my employer. That should be the offense the company is tried for.
Using content to train an LLM is not copying the content. I'm ignoring the silly "but actually" arguments about the content being in RAM so it's "copying". It's using the content to generate a statistical model of token (word-ish) relationships and probabilities. If you write content that is so original in it's wording and I train an LLM against it, then there is certainly the possibility that the LLM could be provoked the recall the exact words you used. You'd have to set the parameters just right to make it happen and I think that proper training would drastically lower if not remove that possible scenario. But even if it doesn't, the LLM doesn't have a copy of that original content. All it has is weights representing those relationship probabilities. Yes, the minutia is more complex, but that is the essence. If my LLM were to generate enough of this essentially verbatim unique content and I tried to publish or copyright it, then I as the user should be on the hook. But then you get into a discussion about how many words in a unique sequence does it take to be infringement?
Obviously, I am not a lawyer.
My summation in all of this is that new laws need to be put into place to handle this stuff because the existing ones are sufficiently non-definitive and/or ill-suited such that every party is forming strong opinions about how old laws apply to new situations and causing massive friction.
Copyright is defined in law and as the original poster stated, whether this is 'copying' as defined by copyright law is legally ambiguous.
Copyright doesn't protect against all forms of duplication. For instance, you own the copyright to your post and grant HN a license to offer copies of it. I have no direct license from you to copy the content of your post; but I can copy it to memory, copy a cache to disk, and make a copy appear on my display.
> For instance, you own the copyright to your post and grant HN a license to offer copies of it.
It’s not a good example, because if you grant a license you give them the right to make copies. The problem is not when Meta got licenses, it’s when they did not.
You can train yourself with every book at any library. You can also train yourself on a large number of movies and TV shows for a small monthly fee.
Where your analogy goes wrong is you're saying you want to "[Circumvent] payment to obtain copyright material for training" to use Workaccount2's words.
So Meta borrowed every book from a library and paid to obtain all of the movies and TV shows? They kept only one copy of every book at any time on their system?
Because I'm certainly not allowed to photocopy a library book in its entirety. And I guarantee you a Netflix subscription doesn't allow me to keep a copy of a movie on my hard drive and use it for training man or machine.
Depends on where you live. In Sweden you can make a few copies of almost anything without violating copyright. There are a few exceptions. Copying entire books was added as an exception in 2005. You can still copy parts of a book. How large parts? I don't know, but I once asked for a copy from a library and they said that a few chapters was fine, so maybe that much (I am not a lawyer).
It depends on your license? I mean strictly speaking if you stream a video you purchase legally over say amazon prime, there's lots of "copying" happening at various levels after those bits leave the data center.
I don't think this is a reasonable argument. I don't think copyright is actually defined in that sense, but is perhaps more focused on consuming the content. Is an http proxy making a copy of something? What about computing an md5 of it as it's streamed through the proxy? Or maybe counting the words in the thing being served in order to track stats? I'd argue none of these fall under copyright, but each is an incremental step towards what it means to train a model.
I'm not a legal expert. My layman's understanding of the case above is Aereo was in violation because they made copies of content - content that the receiver was already allowed to access - available over the Internet to the intended receiver. That is to say, the copying was the problem.
I'm not a lawyer either, but from the summary a key part of the case seems to be that they were distributing the video to people for them to watch.
"Aereo's retransmission of television broadcasts was a "public performance" of the networks' copyrighted work. The Copyright Act of 1976 forbids such performances without the permission of the holder of the copyright. Second Circuit Court of Appeals reversed.
Court membership"
Copyright law does not restrict storing copyright information. It restricts distribution of copyright data without permission. So a computer can store and analyze data but cannot spit it out verbatim. If it spits it out under fair use clause, then it becomes debatable whether the new work is fair use.
Then why folks were arrested for filming in the cinemas? I don't think that's how the law works [1]:
> 106. Exclusive rights in copyrighted works
> Subject to sections 107 through 122, the owner of copyright under this title has the exclusive rights to do and to authorize any of the following:
> (1) to reproduce the copyrighted work in copies or phonorecords;
And later:
> 501. Infringement of copyright
> (a) Anyone who violates any of the exclusive rights of the copyright owner as provided by sections 106 through 122 or of the author as provided in section 106A(a), ..., is an infringer of the copyright or right of the author, as the case may be.
To me it seems clear that Zuckerberg violated author's exclusive right to reproduce copyrighted works. The law doesn't say it is ok to do if nobody knows about it.
For curious, what is considered a "copy":
> “Copies” are material objects, other than phonorecords, in which a work is fixed by any method now known or later developed, and from which the work can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device. The term “copies” includes the material object, other than a phonorecord, in which the work is first fixed.
So an SSD with LLM weights should also be considered a "copy" if from them the work can be "reproduced".
The last excerpt is key: copies are things by which something can be reproduced even with the aid of a machine.
Can you get the LLM to reproduce the work? If yes, that would be a copy.
On the point of what Zuck did, The law states what is prohibited and what exceptions are made against the prohibition, not necessarily what is permitted. If the work cannot be reproduced, then it is not a copy.
As for the cinema part, an argument can be made that the person conspired to reproduce the work. And rule declarations are made that you can’t use video recording device in the theater.
Verbatim is not the point. “Transformative” is. You added nothing to the work when you film a movie at the theater. Google Books (Authors Guild v. Google, 2015) literally shows you photocopies of the pages but is considered transformative and fair use because they offer an easily searchable digital database (and also not harming the book selling market).
I have a weird controversial view on this in terms of how to legally do it, and that is, for your 1 model, you should be only required to buy a digital copy of the work, maybe publishers should make digital copies that are tailored for LLMs to churn through, but then price it at a reasonable rate, and make the format basically perfect for LLMs.
This is actually clever, let the market decide the price and the worth of each book for training. Pricing per model might be tricky, instead annual licensing for training might be better pricing structure. Very quickly all big publishers and big labs might find very precisely what the fair price is to pay per book/catalogue.
It’s weird that you are saying it’s unambiguously illegal. AFAIK, in some cases used for training were initially created by non-profits and transformed sufficiently to strip the copyrights.
I'm not sure if Meta did anything illegal in 2. either.
I thought the copyright infringement was by the people who provided the copyrighted material when they did not have the rights to do so.
I may be wrong on this, but it would seem a reasonable protection for consumers in general. Meta is hardly an average consumer, but I doubt that matters in the case of the law. Having grounds to suspect that the provider did not have the rights might though.
The original complaint alleges that the training process requires copying the material into the model and thus requires consent of the copyright holder. (Copyright protects copying but notably not use, so the complaint has to say they copied it in order to have standing). Then it says they didn't have consent.
They also mention Books3, but they don't appear to actually allege anything against Meta in regards to it and are just providing context.
I don't think it actually changes anything material about this complaint if Meta bought all the books at a bookstore since that also doesn't give you the right to copy the works.
The original complaint is 2 years old though, so I don't really know the current state of argumentation.
Note that incidental copying (i.e. temporary copies made by computers in order to perform otherwise legal actions) is generally legal, so "copying" in the complaint can't refer merely to this and must refer more broadly to the model itself being a copy in order to have standing.
> but it would seem a reasonable protection for consumers in general.
The final say may ultimately come from the Cox vs Record Labels case from 2019 that is still working it's way through the appeal courts.
If the record labels win their appeal, anyone who helped facilitate the infringement can be brought into a lawsuit. The record labels sued Cox for infringement by it's users. It's not out of the question that any ISP that provides Internet connectivity to Facebook could be pulled in for damages.
For Meta these two cases could result in an existential threat to the company, and rightly so because the record labels do not play games. The blood is already in the water.
The source being illegal doesn’t make your use legal. Infact one could argue that it’s equally illegal or worse since a corporation knowingly engaged in illegal activity.
Well obviously, but the converse is also true. The source being illegal doesn't make your use illegal. In the eyes of the law it doesn't matter if something is better or worse, it's the law, it shouldn't be confused with morality. An act is illegal if the law says it is and isn't if it isn't.
Just by being involved as a party does not make you culpable. Murderers are criminals, the murdered, less so.
Choosing to be a party might not make you culpable. You may be an active participant but unaware of the law breaking (being defrauded). Or the law may explicitly state that you can engage with people committing criminal acts and reap the benefits so long as you don't break those laws (or encourage them to be broken) yourself. Some forms of journalism are protected in this way.
Ultimately to have a case you have to state.
1. What law was broken
2. How an action by a party is in violation of that law.
3. That the action actually happened.
The largest problem with this case is not that 3. is in doubt but showing which 1. and 2. they are talking about.
Blizzard managed to get a copyright infringement win against a defendant company that merely accessed their game client (IP) in memory: a cheat reading values of player position
IP that had been previously loaded by Blizzard itself
That link says that 'win' was a summary judgement that was reversed upon appeal.
It seems the case is ongoing, but the case against the company is not one of them committing copyright infringement, but that copyright infringement occurred that they encouraged, enabled, and profited from.
Even if your belief that only the person *providing* the content is liable, do you honestly think a single person found all the content, downloaded it, directly trained the model themselves, and then deleted the content? If at any step the content was given or shared to anyone else for any reason, have they not converted into a provider themselves?
That in itself is a complicated issue, but I would suspect that this does not count as copyright infringement. The entity in possession of the data does not change, the copies and manipulation are performed by employees but at no time do they own what they are manipulating.
If it were true that this constitutes a transfer of possession, then the targets for the lawsuit should be the individual employees, and I don't think anyone wants that.
I would counter that they were operating under the direction of the corporation and, as such, the corporation should be liable. But, I'm not a lawyer.
My overall point is just that a corporation willfully violating copyright should be treated wholly differently than an individual.
I still think that, regardless of any copyright violation by the corporation in acquiring the content, that training with the data doesn't (or shouldn't) somehow taint the resulting model.
>I still think that, regardless of any copyright violation by the corporation in acquiring the content, that training with the data doesn't (or shouldn't) somehow taint the resulting model.
Yes. Building the data set, acquiring the data set, training the model, using the model, and using the output of a model are all separate actions, it is not necessarily the case that any legal standing in one act transfers down the chain.
It's certainly not going to be resolved by a single case.
> AI doesn't actually directly copy the material it trains on, so it's not easy to make this ruling.
IANAL, but it doesn't look that hard. On first glance this is a fair use issue.
What an LLM spits out is pretty clearly transformative use. But the fact that it pulls not only the entirety of the work, but the entirety of MOST works means that the amount is way beyond what could be fair use. Plus it's commercial use. Put it together and all LLMs are way illegal.
> the fact that it pulls not only the entirety of the work
What do you mean by "pulls"?
What matters in traditional fair use is how substantially your output copies the work (among other factors). Your input is generally assumed to be reading/watching/listening to the entire work, and there is no problem with that.
1. Training AI on freely available copyright - Ambiguous legality, not really tested in court. AI doesn't actually directly copy the material it trains on, so it's not easy to make this ruling.
2. Circumventing payment to obtain copyright material for training - Unambiguously illegal.
Meta is charged with doing the latter, but it seems the plaintiffs want to also tie in the former.