"AI training is fair use" isn't settled law: what Bartz and Kadrey decided, and why music is different

"Training AI is fair use" gets repeated as if a court settled it in June 2025. None did. Two judges in the Northern District of California ruled for AI companies on the particular records in front of them. Both rulings were about books, both left big questions open, and one of them reads more like a warning to AI companies than a win.

What did the two rulings actually decide?

US fair use is decided case by case on four factors in 17 U.S.C. § 107: the purpose of the use, the nature of the work, how much was taken, and the effect on the market for the original. Both judges went through all four. They split on the one that matters most.

In Bartz, Judge Alsup found that training Claude on books was fair use, and that scanning print books Anthropic had bought was too. But he held that Anthropic had no right to keep pirated copies in a central library, noting it had pirated over seven million book copies, and ordered a trial on them. The authors also hadn't claimed that Claude's outputs copied their books, partly because Anthropic filtered what the model returned, so the judge looked only at training inputs. That trial never happened: the Authors Guild reports that a court gave final approval to a $1.5 billion settlement, roughly $3,000 a work, on 20 Jul 2026. The settlement releases past conduct on the input side only, not claims about outputs.

In Kadrey, Judge Chhabria gave Meta summary judgment but opened by saying that in most cases, training on copyrighted works without permission will likely be illegal. He said the ruling only showed that these plaintiffs "made the wrong arguments" and failed to build a record, and that it doesn't make Meta's training lawful. The argument they missed was market dilution: a model that can turn out endless similar works in a fraction of the time can crowd the originals out of their market. He thought the effect would hit some kinds of work much harder than others, and he listed songs among the things generative AI could flood the market with.

Why music is a harder case for AI companies

Read together, the rulings tell a rights holder what to prove. Debevoise and Skadden both describe them as narrow, fact-specific decisions about inputs. Music Ally's analysis concludes that rights holders will need detailed evidence that AI models compete in specific markets such as streaming, sync and stock music. Three things make that easier in music than in books.

The purpose is closer to the original. A chatbot trained on novels mostly doesn't sell novels. A music generator trained on recordings sells recordings. The US Copyright Office's training report says almost exactly this: training an audio model on sound recordings to generate new ones serves the same market and the same listener desire, and so is "at best, modestly transformative". It contrasts that with a model built to remove distortion from recordings.

The harm has a mechanism. Streaming pays from shared royalty pools, so every AI track that earns a play takes a slice from everyone else. The Copyright Office says outright that royalty pools can be diluted, cites UMG on "cheap content oversupply", and points to a federal fraud case in which one person allegedly collected more than $10 million in royalties by streaming thousands of AI-generated songs. Reed Smith makes the same pro-rata dilution argument, while noting that no court has yet ruled on whether it counts as legal harm.

There is a licensing market. Both judges dismissed lost training-licence fees as circular. But the Copyright Office says copying expressive works from pirate sources to make competing content is unlikely to be fair use "when licensing is reasonably available", and music companies have since started signing licences. Reed Smith reports that Warner Music Group settled with Suno in November 2025 and moved to a licensed-model partnership.

Lyrics add a fourth point that the book cases never reached: outputs. Music publishers told the Copyright Office that asking Claude for the lyrics to popular songs produced near-identical copies in 500 examples cited in their complaint, and Anthropic agreed in January 2025 to keep guardrails against infringing outputs. We cover that case, Concord v. Anthropic, here, and Kogon v. Google here.

What we think

The June 2025 rulings are a better result for musicians than the headlines suggested. They set a test (show the market harm) that music is unusually well placed to pass, because streaming turns dilution into a number on a royalty statement. Book authors had to argue that AI novels might one day crowd out their sales. A songwriter can point to a pro-rata pool and a flood of uploads.

The winners so far are rights holders with catalogues, data and lawyers: publishers and labels who can document a licensing market and price it. The losers could be independent artists. If the law ends up as "train freely, but pay when you dilute a market", the money goes to whoever can prove the dilution, and licensing deals signed by labels don't automatically pass consent or income down to the people on the recordings. Reed Smith notes that 29 creator and manager groups complained in June 2026 that music companies licensed work without consulting them.

Two practical things for artists, whatever the courts do next:

• Keep evidence. Save royalty statements and streaming data by month. If dilution is the theory that wins, someone will need before-and-after numbers, and collective claims are built from individual records.
• Read your deals for AI terms. Ask your label, distributor and publisher whether they have licensed your catalogue for training, on what terms, and what share reaches you.

When you choose tools, ask what the model was trained on and whether that data was licensed with consent. That question applies to every AI feature in your session, including Veena's. Fair use is a defence a company argues in court after the fact, and "the judge said it was probably fine for books" is a weak basis for a business built on your work.

Disclosure on sources: this article was researched with an AI model made by Anthropic, a party to two of the cases discussed. Every claim above is checked against its cited source.