There's been a couple court cases (that I know of / at least), and one judge was accepting on the argument that model training was a "fair use" while the other was not. I think both of those rulings came down prior to the publication of the U.S. Copyright Office guidelines.

Also, I'm not 100% sure that the U.S. Copyright Office is an authority here. The DOJ and/or Federal judiciary would have the authority to interpret the copyright laws: The DOJ to decide to prosecute, and the judiciary to make binding rulings and/or advise juries. I'm sure both the DOJ and the judiciary will give a lot of weight to the guidelines, but the guidelines aren't actually the law.

In any case, you can read the guidelines and make your own decisions: https://www.copyright.gov/ai/ Part 3 is about training, and I think the damning bits are III, B and D. Part 2 is about outputs, and I think the damning bits are II, B and D.2. (My summaries: 1. Training infringes 2. Outputs that are substantially similar infringe 3. models get no copyright 4. prompts are NOT 'human creative effort' and thus are insufficient to establish copyright 5. human creative effort still gets copyright protections, even when generative AI is used as a tool in the creative process.)

It is likely that commercial generative AI is in violation of a lot of copyrights, yes. Research projects are fair use, but only as long as they stay research projects.

Yup. Give a Markov chain multi-billion parameters and you can get some surprisingly cogent results.

I will freely admit that current LLM architectures include several innovations that make them not actually Markov chains, but it's still statistics and linear algebra. I don't know what thought is, but I'm quite unconvinced that LLMs (or any current generative AI architecture) is doing it.

Alsup ruled in June that Anthropic made fair use of the authors’ work to train

But, Kadrey v. Meta Platforms, Inc. (Judge Chhabria, June 25, 2025) states that “in most cases,” training LLMs on copyrighted works without permission is likely infringing and not fair use and "this ruling does not stand for the proposition that [...] use of copyrighted materials to train its language models is lawful."

The courts are divided, but the copyright office is not.

I think we’d be in a much more agreeable place if we just legally state that a commercial AI tools training isn’t fair use.

Good luck getting any federal law changes through before 2028, at best. So, for at least a couple of years, we get to use the existing laws to address generative "AI"'s use of works still under copyright protection.

Any change in status before then we be policy changes by the U.S. Copyright Office, but they have already come down against training (mostly; the publications are long because there's a lot of nuance).

The document you referenced below appears to be as-yet not officially published

Parts 1 and 2 are officially published.

On May 9, 2025, the Office released a pre-publication version of Part 3 [...] A final version of Part 3 will be published in the future, without any substantive changes expected in the analysis or conclusions.