I don't think this legal analysis is correct. The most actionable copyright problem isn't anything that OpenAI's models do in generating content (whether you consider it transformative or otherwise). The problem is that in creating the model, OpenAI's developers had to copy (into GPU memory, at least) creative works protected by copyright. That copy was not fair use under any of the long-established criteria.

The closest precedent is probably the 9th Circuit's ruling in MAI Systems[1]. Given the Supreme Court's recent ruling against the estate of Andy Warhol, I'd bet on them endorsing the key holdings of that precedent in a hypothetical Silverman v. OpenAI.

[1] https://en.wikipedia.org/wiki/MAI_Systems_Corp._v._Peak_Comp....

Is that the right sense of "copying" though? I thought "distributing" was regarded as a better way to think about the law in this area.

Take a look at the circumstances of MAI Systems. The infringement found was a third-party maintenance engineer copying the user’s licensed software within the user’s own machine for unlicensed use.

Congress later carved out this particular case as fair use. There is no such carve out for copying novels to train LLMs.

Beat me to it.

That said, I'm not sure the caselaw is that clear. For example, in Authors Guild v Google, the courts found Google's use of copyrighted works to build a search index as fair use:

https://en.m.wikipedia.org/wiki/Authors_Guild,_Inc._v._Googl....

In particular, of the four factor test (https://en.m.wikipedia.org/wiki/Fair_use#U.S._fair_use_facto...), in Authors Guild the judge found that the authors didn't experience economic damage from the copying.

In the case of LLMs, can the affected artists argue for economic damages? That's not obvious to me.

Google Books was a wildly distinct use, though. There’s plenty of support in copyright law for fair use claims around indexing and excerpting for such purposes.

As for economic damages, that seems easy: One obvious use of the model is to create derivative works that could economically compete with the original creators’ works.

Simpler examples make this clear: I love Disney animated films. I love them so much I build a model & train it on the complete corpus of Disney animated films. My model is a success, and can produce new films which more or less echo the style, characters, themes of the Disney films.

This is not what Google was doing. The creative work was the well-organized, efficient index, not new books.

I don't think this case is about GPU copying, since that's not what's alleged in the complaint. Mostly, they are alleging that Silverman's book was in a training set that contained an unauthorized copy.

I wrote this post in a way that was meant to be brief, clear, and honestly a little snarky. If I was writing a brief, I'd have gone through a more detailed analysis, but I think fails for a number of reasons, including fair use for the copying, transformative use for the AI system, and frankly a lack of damages because the LLM isn't distributing copies of her work (which is one reason they are trying to establish Silverman as a lead plaintiff in a class action suit). I glossed over a lot of this discussion to keep the post brief. But of course, this post is not legal advice, it's just a quick (and hopefully fun!) blog. I am much more knowledgeable about patent law than copyright law, by the way, and acknowledge any flaws in my post.

Most importantly, I wanted to highlight what I view as a real difference between what copyright holders want from copyright and the policy goals of copyright. What I believe they want is a patent, but what they have is a copyright. They are intended to achieve very different policy goals and as such have different structures.

The fundamental problem:

a) Training the model involved making a non-transient copy of Silverman’s work (e.g. in an encoded form into the GPU).

b) That copy was not a fair use. The purpose of the copy was to train an ML model to generate similar content. OpenAI offers commercial access to the resulting generation mechanism, and indisputably some of what people are paying for is content generation to compete with creators like Silverman.

If OpenAI were only training models for research purposes they’d have a much stronger case for fair use.

I'm pretty sure this doesn't address the key point, that OpenAI and others apparently used a giant dump of pirated epubs as one of their training sources.

Not sure how that would negate the fair use defence.

The question is can you stop it from blatantly offering up a copy of some ones work?

Yes you can create a transformative work but if you can get it to make an exact copy and can't get it to stop doing that I would say there's a problem.

if your a service provider and I send you a DCMA request how do you handle it?

Have you tried to get an AI to create a direct copy of something of any substantial length?

If they're using pirated books, haven't they already committed copyright infringement by keeping these copies? The model itself might be transformative, but I think that is a separate issue.

If the companies had bought all of the ebooks, then they might be in the clear.

But even given that, does that transfer on to the AI model itself?

Or is it "you did a bad thing in the process of making another thing but we can only punish the initial act"

If I watch a pirated copy of Oppenheimer and then summarize the plot to you, the piracy would be a crime, but the summarization would not.

If you watch a pirated copy of Oppenheimer, perfectly memorise it with machine-like accuracy, and then recreate it, then yeah you might be in violation of copyright law.

LLMs are very good at creating things in the style of existing work, but they aren't that great at re-creating existing work verbatim. For example, here's what happens if you prompt Midjourney to draw "the mona lisa": https://i.imgur.com/njOlL5h.jpg

Obviously these are attempts at recreating the Mona Lisa, but none of them are the Mona Lisa. Whether they fall within fair use is an entirely separate question.

It originally being pirated or not would not come into play if its a copyright violation or not. Its a violation if it's 1:1 and billed as such, otherwise its a parody or inspired by

If you did so and as a consequence it was provable that far fewer people saw Oppenheimer, resulting in economic damages to the copyright owner, then the summarization may not qualify as fair use and might very well be a crime.

(Btw I'm not saying that's the case, but that's essentially the argument)

This article is simplistic and wrong. It's not enough to simply transform the copyright material, it's not a panacea. The existing copyright code provides a fair use doctrine that is sufficient for this novel use, and you really only need to change the law if you want to take away copyright protections.

From my previous comment:

Given the potential economic impact on the original artist, the mechanical nature of the transformation and the fact that extracting the style of an artist's work could be deemed substantive, fair use is not in the least bit a slam dunk.

One could argue that use of this sort of AI robs the original artist of the fruits of their labour, neutering copyright protections, and produces works that substitute for the originals, without adding any creative input other than commanding the output from a machine. Given the intent of copyright law, judges would likely be receptive to this argument.

> The AI companies are going to wipe the floor with these litigants using copyright law as their towel because it’s basically impossible to argue that machine learning isn’t transformative use.

The overly aggressive tone of the article is weird from the get go, and it states these (legal) opinions as fact, and never backs them up. From what knowledge or experience does the author draws this from, or is it just the opinion of Generic VC Investor who has investments in affected companies?

The author does get one thing right - copyright law is about making copies. The lawsuits allege that the companies made unauthorised copies of their intellectual property.

Do people have the right to stipulate they don't want their original creative works ingested by an algorithm and then regurgitated out on a prompt thereby benefitting the corporation who owns the algorithm and not the original artist?

I think they do. And whether or not this particular case is won by Silverman is ultimately irrelevant. Society will have its day in court, and eventually this madness of so-called "AI" stealing our labor will be stopped one way or another.

I think it's difficult to argue that people have that right today. As a person, I can read creative works and then synthesize something similar.

For example, if I were a graphic artist, you could commission me to create art "in the style of $ARTIST" and I would be able to legally do so as long as I wasn't copying the source art exactly. You can prompt Midjourney to do something similar, and it will happily comply.

There are obvious ethical questions here, but I'm not sure that would result in a court finding for the copyright holder. We could choose to create new rights which prevent AI from ingesting copyrighted material, but it seems like it would need to be a new law, rather than rights conveyed under existing law.

> As a person, I can read creative works and then synthesize something similar.

Sure, and you'll be sued for copyright infringement if it's too similar. If you've created something new that transforms the original sufficiently and doesn't impinge on the original artists ability to derive benefits from their work, as well as several other restrictions, then you're ok.

A mechanical transformation is an entirely different ballgame. Firstly you'd have to convince a judge that the machine was being creative. Given that the code can't do anything without being fed other people's works, you're then arguing that you're not appropriating other people's work just because you're doing it at scale. Machines using other's creativity (at scale) doesn't mean it has its own creativity. Machines are not human and human creativity isn't a fully understood phenomenon. It's a legal minefield.

I hear you, but at the same time there's nothing in the current copyright law which differentiates fair use afforded to a person from fair use afforded to a machine.

There's definitely an argument that we should have laws which govern machines, but we don't today. That's why I said it would likely need to be a new law that specifically does so.

Maybe they should, but so far as I know currently they don't, at least in the US.

See: https://en.wikipedia.org/wiki/Authors%27_rights

> Do people have the right to stipulate they don't want their original creative works ingested by an algorithm and then regurgitated out on a prompt thereby benefitting the corporation who owns the algorithm and not the original artist?

Most times somebody makes this case it's explicitly framed as "AI owned by big corporations". I can see why this makes the argument more compelling but does the argument stand without this?

i.e. if we are talking about an open model with no corporate control then do you feel your point is as strong as before? These models indisputably exist. I think there is still a reasonable case to be made but I think it's slightly dishonest to make it in this way when we know that things have moved on.

Congress can't extend the coverage of copyright without limit. The Copyright Clause in the Constitution limits Congress. This is why the US doesn't have database copyright or "sweat of the brow" copyright. EU law is different, though. The EU does recognize database copyright.

Machine learning tables of weights are probably not copyrightable, but we can't be sure until there is more litigation.

Overzealous copyright protection is what you might call a "first-world problem". Every time a new technology comes along, the copyright defenders get nervous. These people believe that society owes them and owes their ancestors a living into perpetuity. Shakespeare never got paid what he was worth but two-bit comedians like Silverman won't let that deter them from demanding more.

Yet another chapter of the dementia surrounding the current copyright laws. We don't tax streets, we don't tax public spaces (mostly), but still western societies keep pushing forward the idea that carrying devices in our pockets capable of bearing almost all the available human knowledge (that can be almost instantly cloned), and still selling individual files is a logical concept.