I'd say it is possible to produce exact data as well. Try "Provide quote from King James' Bible Genesis :1-25" with chatgpt. You'll get a verbatim text. You can get the same with things like Moby Dick, but when I typed "Provide the first five sentences of the book A Game Of Thrones" I got:

Certainly! Here are the first five sentences from the book "A Game of Thrones" by George R.R. Martin:

"We should start back," Gared

This content may violate our content policy or terms of use. If you believe this to be in error, please submit your feedback — your input will aid our research in this area.

The model is clearly capable of reproducing verbatim data I think.

It's still surreal that this is considered Fair Use, and even defended relatively recently (2013). It's hard to say where the ruling will land ultimately, but there seems to be an argument that verbatim reproduction doesn't matter.

Siskel and Ebert didn’t need to pay rights holders to extract from their works for public criticism.

Because 'transformative' is a pretty dangerous word to use in this context.

I strongly feel that this is a terrible metric for comments on the internet.

First, the person you’re replying to has nothing to gain and a lot to lose by saying "yes".

Second, it invites silly corner case nitpicking. Their comment is written in reasonable plain English for other users reading plain English. It’s not a legal contract, and so leaves lots of loopholes. Sure, you could create a likely non-transformative LLM by training it on nothing but the text of Harry Potter with fitness measured by how accurately it exactly reproduces the complete text of Harry Potter, but that’s not what reasonable people are doing with LLMs.

In a legal context certain words have immense power. In the context of copyright 'transformative' is one such case. It's a very fine line between 'transformative' and 'derivative' and you don't get to preempt the judiciary about how they will see things.

Designing systems around what people should do, as opposed to what they actually do, has proven time and again not to work particularly well in practice. I'm sure you've seen countless examples of how people track paths through manicured grass fields. The landscaper will complain about how people should walk and they'll put up signs to no avail.

The fact is, we (including me, BTW) are frequently wrong about a lot of things, and when there's little riding on it, we can ignore that most of the time. With subjects like medicine and law, however, where a mistake can cost you your life or lots of money, we want to make sure people are getting the best advice possible. That's why we require licenses to practice medicine and law, and we have governing and ethics bodies to regulate how professionals operate their practices.

Correct, so people should (and do) go to the people who have these licenses, not random people on the internet. I don't even understand what your solution, or even problem, is. It seems like you're suggesting that everyone, whenever they speak on the internet about anything vaguely related to medicine, law, or hell, even regulated fields like engineering, should disclaim that they are not speaking in such a context. And I saw that that is a ludicrous task that is expected of one to do. So if you have any better solutions, let me know.

This isn't how "fair use" works, in the sense that there can never be a blanket assurance like that. Also, whether the result is "transformative" is just one of many factors (see audio sampling/remixing).

“The Godfather” film is absolutely a transformative interpretation of Mario Puzo’s book and a fully distinct, valuable work of original art. Paramount still needed to pay Puzo for the right to base it on his words.

What if you train it only on my huge repo of GPL code? You are just remixing my code.

Now you maybe think "let me train on 2 different devs GPL code", the remixed code will probably be 50-50 and you can get away with it ?

If the 2 number is too small then tell me what the number N should be ? From how many people you need to "steal" code , mix it and the output is "original" ?

Edit: my opinion is that AI should be fair, if you train it on open source then model should also be open source and output should also be open source.

This is all to say: the question about copyright remains exactly the same regardless of license.

I do not disagree with what you said, only that in reality, this is not the only way business conflicts are decided.

This never happens, you will first learn from a book or tutorials.

But your idea is sound, have Microsoft buy books from the authors and train the LLM on those books then have the LLM solve new problems. If is an AI and not a text interpolating tool then it should be able to learn like humans from a few books.

I have learned to code almost entirely from reading code. I hate tutorials and books.

I occasionally skim docs but mainly for the code examples.

So "never happens" is a curious take.

The word "remixing" here is useful because it will fit any conclusion the reader prefers.

Arguably even in your reductive example, the result would be non-infringing. Or not. Which conclusion you reach is exactly the topic under debate. Isn't this textbook question begging?

Imagine I get the Windows source code and rename the variables by adding a "314" after each varaible, after each function name and rebuild Windows, in your definition this is remixing and fair ?

Where you like it or not this is an undecided area both legally and morally. Pretending it's clear cut is either disingenuous or delusional.

My simple example is to show that is not as simple as "the AI earned from N devs GPL code and now it can spit new original code without ZERO concerns", we know how this stuff works and that it can spit out the exact training input in some cases.

So IMO a judge should ask the question "from how many people you need to steal, mix the input to be sure the output is actually original".

And about the thing "if I read someone code it is not stealing" , hyumans are different and even for humans it is not allowed to read the code of your competitor and then write new code using that knowledge.

The economic part of copyright is transferable in the EU just as it is in the US, only certain moral rights (such as the right to attribution) are inalienable.

In other words, law isn’t a programming language.

It can go either way.

And why would it matter?

Except that “fair use” is mostly an American thing. In many other jurisdictions (especially those with of civil law) there's such a wide principle, and there's only specific laws allowing some explicit kinds of use of copyrighted material that the law allows. In those jurisdiction, most uses of generative AI trained on copyrighted material are, more likely than not, illegal at least until the legislator actually changes the law.

Have you read the recent SCOTUS decision in Warhol v Goldsmith? Because that's a pretty major redefinition of transformative for the purposes of fair use, and not in a good way for arguing that generative AI is fair use, especially because it ties transformative to the market impact. That generative AI is generally creating outputs that are directly competing with inputs (particularly in the case of generating images, where it's clearly competing with stock images) would make it dramatically less likely that a court would find that it is in fact transformative.

The benefit that generative AI has is that, when claiming copyright infringement, you need to specify individual works that were infringed. It's not enough to say "this work is an amalgam of these other ten thousand works, and we can't really tell you how."

I could imagine if generative AI gives an identical, word-for-word match for an individual piece of source material it could be in trouble, but that's also the easiest type of thing to prevent from an AI company perspective.

The fact is that existing copyright law just can't really encompass the kinds of societal concerns we have around generative AI.

At any rate you can force the infringer to disclose what works they use as input.

Copyright law doesn't encompass novel uses, but courts can and will deal with it.

Some cases are pretty obvious, but even literal copying isn’t always copyright infringement (e.g., if the material is arguably not eligible for copyright protection).

That's a little bit like "If a tree falls in the forest but nobody hears it..."

I mean, sure, "theoretically" any number of things can be infringement. But it's obviously a gray area, so it only really matters when somebody brings a suit and a work is found to be legally infringing.

Purely mechanical modifications may not be considered transformative, and there's an argument to be made that LLMs are purely mechanical (in fact a US district court recently ruled that AIs cannot be authors of copyrighted works).

I thought that was because only humans and other legal persons can legally author things, not because of anything subtler about the nature of LLMs. See also the case where the monkey managed to take photos of itself. I'm not a lawyer, though.

It very explicitly was and made a point of noting that it was not addressing anything about whether and when a human author could hold a copyright on a work authored using AI.

Big bet on legal costs based on something being "likely".

Just because Copilot might be itself a transformative work which is itself allowed to exist, that doesn't at all necessitate a conclusion that the developers who are using it are going to or should somehow be guaranteed not be committing their own copyright sins if they try to incorporate its output into their own works (any more so than one can or should assume all of the outputs of another human being are free of copyright entanglements, even though no one is as-yet claiming a human being is themselves infringement just because they saw another work).

Everybody seems to be saying this, but I really don't think there's even 50% chance of it happening.

Google books was fair use because it was a public benefit and did not take away from publishers or authors, to the contrary it helped people find their works.

Compare generative AI which extracts the essence of people's works and recreates similar works (in terms of style, etc) while cutting out the original authors completely. This potentially denies them the fruits of their labor. It's notable that it's a purely mechanical process and no human creativity is involved, except that which is extracted from other authors. Mere prompts don't count.

The argument you're suggesting will hold is essentially "yes we're using copyrighted works, but we're doing it at scale and blending it, so that's ok".

Commoditized goods allows the bad to be sorted in with the good, allowing a price to be put on the commodity. Great where it's applicable but horrendous when it's improperly done - ie, home loans, or intellectual property.

If your commodity markets aren't properly regulated you get a race to the bottom. If you are trying to commoditize something that shouldn't be, it's effectively enables white-collar looting or money laundering.

LLMs are typically implemented in a way that makes them non-deterministic (i.e. temperature > 0).

Second, the way we've seen generative AI be used is not really the same as it was touted originally, that a mere prompt could replace an entire artist's work. A year later, we see that most people, artists included, don't use it as a verbatim text to image machine, they use it as a tool. See apps like ComfyUI or others which allow Node based or layer based image creation and editing, which even Photoshop now has. It's the same as Copilot and ChatGPT, it's not replacing any programmers, just increasing their productivity Given that, it is not looking like generative AI is hurting one's professions, quite the opposite.

I don’t think any but the most copyleft segment of society thinks it would be reasonable for a generative AI trained on exactly one persons work to be used for profit by someone else.

An AI trained on two or ten people’s work probably feels the same for most folks, but what about when it’s thousands or millions? What if instead of one persons work it is the works held in copyright by an entity like Getty Images?

Why do you think that? It doesn't seem obvious to me at all.

So I think the key part is "used for profit".

Wasn't suggesting it is. The point is that the tool is used to create things that substitute for the original authors' work by ingesting the works of those authors. The impact of the copying matters when weighing fair use.

If I use your copyrighted works to supplant you in some way, even as a part of a large group, then it's unlikely to be deemed fair use.

Meanwhile, drawing Mickey ears on the wall of a kindergarten is not safe.

If you feel strongly that generational ML somehow launders copyright out of the bits, train an image generator purely on Disney copyrighted material and share the model on the web, see how well that works out.

https://civitai.com/models/24/modern-disney

That's true it's probably 99% plus it happening or at-least that's the conclusion that the experts and lawyers hired to help evaluate AI startup valuations are coming too. Hired by banks, venture funds, short selling shops, etc plenty of people who don't depending on it being ok to make money.

> "yes we're using copyrighted works, but we're doing it at scale and blending it, so that's ok"

I mean you know collages are legal right? You literally take 100s of copyrighted pictures and put them together and suddenly it's perfectly legal and ok.

Contrast LLM-created code which is certainly a substitute for the original copyrighted work.

Really? When has this been done?

Only if it’s sufficiently transformative. There was recently a case that hit the US Supreme Court about this subject regarding an Andy Warhol adaptation of a portrait of Prince [1]. So, in the US, fair use in this regard requires some amount of substantive transformation of the material. But, as we are talking about AI algorithms, there isn’t a person in between the model and the training data. The argument here is whether or not a person is required to make a transformative use of the material (and thus fair use applies). Given that AI generated (and non-human animal generated) works aren’t copyrightable due to the lack of human involvement, I’d wager that any AI use of copyrighted material won’t get fair use protections.

[1] https://www.eff.org/deeplinks/2023/05/what-supreme-courts-de...

What are the odds the market leaders in LLM right now are just the current day version of Borland-style compilers before open source takes it over?

I've heard arguments the infrastructure part is a long term barrier to entry for OSS development, which will continue to remain in the future. But I don't know enough about it.

Who knows maybe the legal/gov world will move slow enough to miss the bulk of the money-extraction opportunities before OSS takes over and the reality of this problem never going away fully kicks in.

Open source models would need benefactors with deep pockets.

Only if you ask it to. At which point the person asking is at the very least culpable as well of violating someone's IP.

It is also illegal for me to pay someone to write Micky Mouse fan fiction (though if I don't publish it, this gets more murky).

> The argument you're suggesting will hold is essentially "yes we're using copyrighted works, but we're doing it at scale and blending it, so that's ok".

I want to flip this on its head: the argument you are suggesting is essentially "LLMs should be illegal because they can be asked to break copyright at scale!" It isn't illegal to be an author for hire, even though someone could potentially ask you to write fan fiction for their personal collection in the style of Tolkien, but because an LLM can do it at scale, it is illegal?

Anyone could use those tools to download creative common files and linux ISO, but those arguments did not succeed in the legal system. Bittorent as a technology was however not made illegal, as could be seen in games using it to distribute patches.

Feature extraction is literally a form of lossy compression. You can prod DALEE to make obvious copies of some of the works it was trained on, but even seemingly novel images could contain enough similarities to training material to be problematic.

https://www.heswithjesus.com/tech/exploringai/index.html

I’ve also seen GPT spit out proprietary content word for word that’s not licensed for commercial use that I’m aware of. They probably got it from web crawling without checking licenses.

What I want more than anything in this space right now are two models: all public domain books (eg Gutenberg); permissive code in at least Python, JavaScript, HTML/CSS, C, C++, and ASM’s. One layered on the other but released individually. We can keep using that to generate everything from synthetic data to code to revenue-producing deliverables. All with nearly zero, legal risk.

So either we carve out an explicit exception that machines aren’t allowed to do things that are remarkably similar to what humans do… which would be a massive setback for AI in the US.

Or we agree that generative models are subject to the same rules that humans are — they can’t commit copyright infringement, but are able to appropriately consume copyrighted material that a human would be able to consume.

The second option seems to me to be much simpler, nicer, and more appropriate than the first.

The standard isn’t “I think that looks like an Andy Warhol picture”, it is “That is substantially similar to a specific Andy Warhol piece”. Copyright doesn’t protect style.

> Feature extraction is literally a form of lossy compression.

This is one way think of neural nets, another is that they find the topological space of pictures.

But these are just models of computation, which aren’t especially relevant in the same way that it isn’t relevant what produces an infringing image, just that it is produced.

Which brings me back to my original point: there are a two different barriers for generative ai: is the model itself transformative, and is the primary purpose of the model to generate copyright infringing material.

With respect to the first… I have no idea how someone could argue that the model itself isn’t transformative enough. It isn’t “substantially similar” to any of the works that it is trained on. It might be able to generate things that are “substantially similar”, but the model itself isn’t… it’s just a bunch of numbers and code.

Regarding the second: I have less experience with image models, but I use chatgpt regularly without even trying to violate copyright, and I don’t think I’m alone, so I doubt you could make an argument that llms have a primary purpose of committing copyright infringement.

That really doesn’t fly legally because any digital format is ‘just’ numbers.

But I think the greater point still stands. In order to call the model itself a copyright violation you would have to say it is "substantially similar" to thousands of original works. Then in order to make it illegal you would have to say it wasn't "transformative enough" to be considered fair use.

I can't come up with an argument for either one of those points that holds any water at all.

This music industry has been going through exactly this for the last few years and the courts have recognized that the creative process necessarily involves copying and that a small amount of copying is not infringement.

Critically it’s not just a question of what percentage of a work is a copy of the original but what of the original work was copied. IE copying 3 lines in a book is a tiny fraction of the book but if you coped half the poem it’s well past the de minimis threshold.

Similarly only a small percentage of a giant library of MP3’s comes from any one work, but that’s not relevant.

Copilot is taking things like “reverse a string” or “escape HTML tags”, that have very little originality to start with. This kind of common language is analogous to the musical motifs that have been also found to be under the threshold.

Where generative AI ingests copyrighted works in order to work and bases its output on it, then it is copyright infringement, equivalent to 'straight piracy' of all that it ingested, unless it's deemed fair use.

What Google does with its search engine, for example, is fair use, what Napster did was not.

This is an absurd standard. Is it copyright infringement when a human "ingests" copyrighted work and bases their output on it? Because that's commonly called inspiration and is how every artist creates their work - through experiencing other works and using that cumulative inspiration to form their own product.

Copyright infringement is already ridiculously restrictive as it is, this proposal not only fundamentally misunderstands how generative AI works but penalises AI for doing what humans do everyday.

Learning, by human or machine, means extracting a copy of the essence of something and yes, storing that essence in a lossy way. It seems like learning from copyright-encumbered material ought to either be illegal for both, or legal. I know which world I would rather live in.

* Does the model itself violate copyright? * Does the output of the model violate copyright?

I don't know how you could make an argument that the ingestion of information into a model through a training procedure in order to create something that can generate truly unique outputs isn't transformative of the original works. The legal standard for a new work to be considered a copyright violation of an original work is "substantial similarity". I don't know how you can make an argument that a generative model is "substantially similar" to thousands of original works...

Honestly, I'm not even sure if "fair use" comes into play for the model itself. In order for fair use to come into play, the model has to be deemed to be violating some copyright. Only once it is found to be violating does "fair use" come into play in order to figure out if it is illegal or not.

The second question is the one where fair use is likely to come into play more. And this question has to be asked of each output. The model's legality only becomes an issue here if, like Napster, you can't argue that the model has much point other than violating copyright. Napster didn't violate copyright (the code for Napster wasn't infringing on anything), but it enabled the violation of copyright and didn't have much point other than that.

I don't think you can make that argument though. I use ChatGPT most days, and I've never gotten copyrighted material out of it. I could ask it to write me some Disney fan fiction, which would violate a copyright. And I think there is a valid legal question here about who is responsible for preventing me from doing that. This is where I think the gray area is.

This isn't necessarily true. It's entirely possible for a model to regurgitate a chunk of GPL'd code without you knowing that's what it's done.

Code is also tricky: there are a finite number of ways to write an algorithm, and I’m sure both that multiple people have written the same version of left pad for example, and that it is not possible to copyright something small like that. When the code gets bigger, the likelihood of an llm spitting out large chunks of GPL’d code seems vanishingly small (without asking for something specific like that). Though I’d love to see examples to the contrary.

Though that is definitely a simpler prompt than I would have expected was necessary to get such a result. Thanks!

(The first example also isn’t the same code. It is very close, and definitely similar in style, but it isn’t clear that code would a)run, or b) would work as expected. I need to sleep though, so I’m not sure how much that matters.)

Are there any crawlers used for commercial purposes which refuse to remove sites from an index if they ask? The distinction from OpenAI is that there is no way to be removed from openai's training set.

You can remove yourself from the crawler not but not from what they previously crawled.

There’s no law against “using” copyrighted works, there is a law against copying and distributing them.

Fair use analysis doesn’t come into play unless we’re dealing with clearly established copyright infringement. What LLMs do doesn’t clearly qualify as any of the behaviors reserved to copyright owners. For example, it certainly doesn’t “copy” the things it’s trained on by any legal definition.

Law works on precedent and analogy when there’s no clearly on-point statutes or case law. The most analogous situation to what transformer models do is a person learning from experience and creating their own work _influenced_ by what they’ve observed. That behavior is not copyright infringement by any stretch of the imagination. The fact that it’s done with a computer is not as important as people seem to think it is.

What about pictures still containing watermarks? Regardless of the actual legality, this does not fit "certainly".

> The most analogous situation to what transformer models do is a person learning from experience and creating their own work _influenced_ by what they’ve observed

No, it is not. It is called machine "learning" so clearly that is a fly made out of butter. Maybe courts will agree, maybe they won't, but the analogy to human learning is quite strenuous at best.

https://www.copyright.gov/title17/92chap1.html#106A

The most closely applicable existing law is that of “derivative works” but those require human authorship, so it’s far from clear that those would apply to AI output either. Ultimately this is going to be hashed out in the courts until some actual laws are written to deal with it.

(IANAL)

It's taking verbatim digital copies and using a form of lossy compression to transform them, which I think is clear when looking at things like auto-encoders.

That’s inconvenient for opponents of this technology because they would prefer to ban the training itself, but there’s not a good justification under existing law to do this.

"I'll keep saying it every time this comes up. I LOVE being told by techbros that a human painstaking studying one thing at a time, and not memorizing verbatin but rather taking away the core concept, is exactly the same type of "learning" that a model does when it takes in millions of things at once and can spit out copyrighted code verbatim."

(I also love it when they're deliberately obtuse about it too. The past decade has made me sick of this trolling tactic.)

If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.

p.s. Also, please don't copy/paste comments on HN.

Even Microsoft is couching their guarantee here with an exception for this very case.

https://www.notion.so/DSM-Directive-Implementation-Tracker-3...

https://eur-lex.europa.eu/eli/dir/2019/790/oj

The TDM4 copyright exception allows datasets to be created consisting of copyrighted works, as long as there is a mechanism for rightsholders to opt out. This seems like the best of both worlds: the dataset is transparent, rightsholders can assert their rights, and certain AI companies can train on copyrighted material.

Of course, this doesn't grant commercial rights for the trained model, only scientific and academic research rights. (I.e. it's fine for Meta to train and release a LLaMA model trained on books, as long as they're not commercially profiting from it, and there's a mechanism for authors to opt out.)

I'm talking with Jordan from https://spawning.ai to try to build some kind of opt out system that makes sense for books. One could imagine doing this for music too.

This is a European law, but unlike other overreaching EU regulations, this one seems like an extremely sensible compromise.

I really don't like this--opt-out never works because the scale advantages are backwards. It places the burden in the wrong place. The aggregators should have to get opt-in.

Look at YouTube. Because of "opt-out", lots of people monetize content that they have no right to and it's up to the original author to have to fight the scale of a zillion uploaders. Only the biggest entities can do that.

YouTube (and everybody else) should have to assert "You, the uploader, own this content" when they ingest it. Nothing else works.

I wouldn't mind an exemption for research use, though.

This is a fairly big deal since right now there’s no incentive for AI companies to disclose their training data, and it seems unlikely that legislation to that effect will be enacted anytime soon. Whereas this opt out mechanism is already getting widespread adoption in the EU.

>Look at YouTube. Because of "opt-out", lots of people monetize content that they have no right to and it's up to the original author to have to fight the scale of a zillion uploaders. Only the biggest entities can do that.

>YouTube (and everybody else) should have to assert "You, the uploader, own this content" when they ingest it. Nothing else works.

Thankfully, in a rare turn of fate, capital will be on the side of the laissez-faire instead of the stringent anti-copyright-infringers for once. You do not own the rights to material created by a generative AI.

Microsoft has tried to get copyright infringement claims based on CoPilot dismissed and so far it has failed.

https://www.courtlistener.com/docket/65669506/147/doe-1-v-gi...

Arguably it is only Microsoft and OpenAI who know the full extent of copyright infringement, if any, occurring through the use of "AI". If one of these cases like Doe v Microsoft gets to discovery then this maybe knowledge will no longer be their kept secret. Perhaps the floodgates open.

Who knows. No one can predict the outcome.