@cwebber Fair use is complicated enough that I don’t think we can generalize. Not from the initial ruling that AI companies are doing fair use, nor from this ruling that the AI company is not doing fair use. This ruling seems to have a factor where the AI product is directly competing against the original copyright holder; I suspect in other fair use cases the court will reach a different opinion. This is of course very interesting to watch from the sidelines.

@cwebber I don't think decompilation is a necessary step at all. For an LLM, object code is just as easy to "read" and crib off of as source code.

It's only human brains which have more difficulty reading object code.

So, let's assume Claude and the others are trained on tons of copyrighted object code. Everything they can get their hands on. (This is a super safe assumption.)

This means that EVERY bit of AI generated code is tainted by copyright violations outside of fair use. If you use AI

@cwebber this is a weird situation in which big corp endorses the AI hype yet also shoots themselves in the foot (sueing others for crimes they are guilty of themselves too).
Safest bet for FOSS is to stay clean as much as possible indeed.

@cwebber There's also a different approach to dealing with genAI infringement that depends only on the output the LLM produced. If you show that the output is legally close enough to be considered identical (allowing for scène à faire and such) and that the defendant (or the LLM) had access to your work, whether the training was infringing or not doesn't matter. The burden is then on the defendant to prove they didn't copy the work.

This is the same as with humans. Someone can learn to...

@cwebber It's interesting how this is happening now, and not a couple of years ago when AI firms were frantically despining every book in existence to build their eliminate-staffing-overhead machines.

It's almost like copyright law isn't about everyone's rights at all.

(Not that I oppose this particular development. It just should have been wrong all along, for everybody doing it.)

@cwebber

Hmm, now I kinda want to buy some tokens and hook it up with a few Mario game Roms and let it rebuild it. Then drop it anonymously somewhere "by accident" or something.

@cwebber Note that this case was an LLM that could replace lawyers. I'd expect them to fight back.

@cwebber And this is just US law. I've already warned that European copyright law is much stricter:

There's no fair use. Every exception is a law.

New ways of using a work require explicit new licenses. The old license did only permit what the licensor said you can do. This doesn't include things not foreseen. So even a BSDL style license doesn't allow you to train LLMs with.

So maybe you can train LLMs with pretty recent stuff licensed under BSDL, because now that's something the author would expect, but you can't take all the old stuff out there.

@cwebber The japanese will invade?

@cwebber
Just earlier today saw a video with the recompilation stuff happening in video games the last few weeks has led people to merge games, they showed Mario running around inside Halo shooting mobs and a Mii character bowling at enemies in a shooter, so yeah, the bear has officially been poked, let's see what do they do

@cwebber "Vibewashing" a program's copyright away. Is that original with you Christine? I nominate that for new word of the year. OED take note.

@cwebber as soon as they touch Nintendo's property, AI is going to be as good as dead 😄

@cwebber @jlink @mttaggart this is going to be the least satisfying “I told you so” of my entire life

@cwebber@social.coop

> The third circuit issued a recent ruling in the opposite direction of the previously well cited federal ruling, that AI training on copyrighted material is NOT fair use: https://www.courthousenews.com/ai-training-of-copyrighted-material-not-fair-use-third-circuit/

It seems hard to generalise that ruling; training a model to produce headnotes and specifically headnotes on headnotes you acquired by inducing a third party to commit copyright infringement(!) is not really comparable to training an everything model on everything.

In fact the ruling takes pains to declare its non-applicability. The judgement starts with:

> Under ROSS’s framing, this case appears to concern the
> future of AI legal technology. But appearances can be deceiv-
> ing. In truth, this is no more than an ordinary copyright case;
> we must determine whether Thomson Reuters’s materials are
> sufficiently original for copyright protection and whether
> ROSS’s use of those materials was fair. The undisputed evi-
> dence reflects that Thomson Reuters’s materials possess a cre-
> ative spark, and ROSS aspired to be a direct competitor by us-
> ing them for a highly similar purpose. We therefore hold that
> Thomson Reuters’s materials are copyrightable and that
> ROSS’s use was not fair. Thus, we will affirm the District
> Court’s order granting Thomson Reuters partial summary
> judgment.

and possess the footnote:

> 7 We note that the United States Department of Justice (“DOJ”)
> filed a statement of interest relating to generative AI in a pend-
> ing case in the Southern District of New York. Statement of
> Interest of the United States, In re: OpenAI, Inc. Copyright In-
> fringement Litig., No. 1:25-md-3143 (S.D.N.Y. Sept. 1, 2026)
> (hereinafter “DOJ Statement __”). The concerns raised in that
> separate case do not apply here. The DOJ relied on Bartz v.
> Anthropic PBC, 787 F. Supp. 3d 1007, 1014 (N.D. Cal. 2025)
> to contend that training an AI large language model, which can
> “generate original responses,” is a transformative use. DOJ
> Statement at 10; see also Kadrey v. Meta Platforms, Inc., 788
> F. Supp. 3d 1026, 1034 (N.D. Cal. 2025) (exploring whether
> generative AI programs transformatively use training materi-
> als). Unlike the AI models in Bartz and In re: OpenAI, ROSS’s
> AI platform cannot generate original expression, and the evi-
> dence here supports the opposite conclusion about transforma-
> tiveness. The DOJ also argued that the large language model
> training in OpenAI did not result in “substitutive competition.”
> DOJ Statement at 13 (citation modified). Here, ROSS trained
> its AI for the purpose of creating a commercial substitute for
> Westlaw. In any event, the DOJ’s filing in In re: OpenAI
> demonstrates that the DOJ is tracking these issues and knows
>how to assert its interests, but the DOJ notably did not do so
> here.

@cwebber it looks like we may not be waiting very long -- ArtCraft, the genAI toolkit, has launched a straight up clone of Adobe's suite, with the same UI, shortcuts, file format support and everything. But it's "clean room," because nobody at ArtCraft has decompiled any Adobe software at all! Riiight? https://arstechnica.com/ai/2026/10/software-is-over-bold-ai-developer-takes-aim-at-adobe-with-open-source-clones/

@cwebber by the way that also means that pieces that have no copyright at all may not be used for llm training. At least the rule in many countries is that if you created something, put it out for free on the internet, does *not* mean you can do anything with it, such as reselling. You automatically have that copyright and anyone who wants to use it for their business must ask you.

@cwebber Companies selling AI, getting abused by AI users: Leopards eating *my* face?

@cwebber holy shit that's huge

@cwebber orbital lawyer strike incoming 💣

@cwebber Video game "vibewashing" is very active in practice. (Modern Vintage Gamer did a video recently on that: https://youtu.be/mYZPHdKKMaI).

A quick look on Quiver (https://quiverlauncher.com/) reveals quite a few recomps of Nintendo games, including big franchises like Mario and Zelda.

That these aren't getting lawyer-bombed into oblivion is quite revealing in my opinion, as if it was somewhat in the interest of by corporations that "vibewashing" remains legal.

[...]

@cwebber Problem is copyright protections will only (practically) be applied by big corps against independent creators.

When big corps want to train on independent material from independent creators, they'll be given a free pass.

@cwebber I don't think the 3rd Circuit's reasoning is going to hold up, mostly because it would imply that natural intelligence using someone else's ideas to start their own business would be an infringement of the work. That it can be 'true' in movie scripts and 'false' in fast food stores is an issue. The *only* good thing might be to clean up US copyright law, but also unlikely.

@cwebber I've heard inconsistent claims that the last one is already going on

@cwebber it also drives me crazy how parochial my fellow Americans are. Even if they did decide, at the federal level, once and for all what the status is of output in the US, that's not applicable elsewhere. And we're in an international environment where trust between national leaders is low and some sort of treaty on this seems unlikely.

@cwebber <<GAME OVER>>

(or at least very very expensive!)

@datarama yes, it is their MO