The DOJ says AI training is fair use. Music's best arguments were never about training

If you make music, a 20-page filing in a newspaper's lawsuit is worth your time. On 1 Sep 2026 the US Justice Department filed a statement of interest in the consolidated OpenAI copyright cases in Manhattan, asking the court to treat the training of large language models on copyrighted works as fair use. The case is about news articles. But the argument the filing attacks hardest is the one the music industry has leaned on most.

What did the Justice Department argue?

The filing (Doc. 316 in case 25-md-3143) was filed in the name of Associate Attorney General Stanley Woodward Jr. Its main points:

• Training is transformative. Copying works to train a model is "exceedingly transformative", because the model learns statistical patterns rather than reproducing the expression.
• No cognisable market harm. Training reveals nothing to the public, so, the DOJ says, it can't be a substitute for the original. Outputs that compete in the same genre without being substantially similar don't count either.
• Licensing is a barrier. Paying for training data would favour the largest tech companies, and the fees would disproportionately benefit legacy media. The filing also frames this as national security: restrictive rules would hand an advantage to foreign rivals.
• Wide reach. A footnote says the arguments apply to every party in the litigation and the related cases, including book authors and publishers.

Bloomberg Law reports that a Times spokesperson, Graham James, accused the administration of siding with "a handful of trillion-dollar AI companies". NOTUS quotes Brandon Butler of the Re:Create coalition calling it a big change from previous administrations. Adam Eisgrau of Chamber of Progress told NOTUS he couldn't recall another government filing that stated a view on fair use so openly. Copyright lawyer Jonathan Band called the positions "very mainstream". A statement of interest isn't a ruling, and the court still decides.

Why music should read the footnotes

It goes after the dilution theory. In May 2025 the US Copyright Office's report on generative AI training said the fourth fair-use factor "should not be read so narrowly". It argued that AI's speed and scale risk diluting markets for works of the same kind as the training data. Its examples of royalty dilution came from music: a Universal Music comment about cheap generic content diluting human creators' royalties, an indictment over more than $10 million in royalties from fraudulently streamed AI-generated songs, and the 2024 label complaint against Suno. The DOJ calls the same theory, as adopted in the 2025 Kadrey v. Meta ruling, "deeply flawed". It says the Register of Copyrights, who it notes is challenging her removal, gave "threadbare" reasoning that deserves no deference.

That matters because dilution describes music's problem better than any one infringing track does. A flood of competent, generic tracks pulls listeners and royalties away from the people whose recordings trained the models, without copying any single song closely enough to sue over.

It treats licence deals as beside the point. Footnote 13 says publishers can and do sign licensing deals whether or not training is fair use. So a licensing market that exists doesn't prove a licence was legally required.

It's narrower than the headlines. The filing is about training on written works. It describes the acquisition stage, then limits itself to "the training stage", so it doesn't address how data was obtained. It also concedes that at the output stage, uses "may not be transformative" if a model reconstructs an original work. Separately, it notes that the President has encouraged Congress to consider a federal framework protecting people from unauthorised commercial use of AI digital replicas of their identifiable attributes.

What we think

The DOJ's brief maps where music's AI fight now has to happen. A case built only on training is uphill in the US while the federal government argues the other side. The claims the brief leaves open are outputs, how the data was obtained, and identity, and those are where music is strongest.

There's already a model for the output argument. On 31 July a Munich court ruled for GEMA against Suno over six songs. According to the court's press release, it found the songs reproducibly contained in two Suno models. It also applied US fair use to Suno's US training copies and rejected it, because the works showed up in the outputs. The DOJ says outputs shouldn't bear on training. The German court said a model that can play the song back has copied it. That link is easier to prove with a melody than with a news article, because a melody is short, distinctive and recognisable.

Identity is the other route, and it belongs to artists rather than catalogue owners. A day before the DOJ filed, Jason Isbell and three other musicians sued Suno under publicity and biometric privacy laws, not copyright. Their complaint says the claim "would be identical" even if every recording Suno ingested were in the public domain. If training becomes free, a performer who signed away their masters may have no copyright claim at all, but they still have a voice and a name.

We're critical of one part of the DOJ's reasoning. It frames licence fees as subsidies for legacy media and ignores that the dilution it dismisses lands on working musicians first. It is fair on one point, though: a licence-only world can favour whoever owns the most catalogue, and the people who wrote and played the music don't automatically see that money.

For artists, the practical steps are the same either way. Register your works. If a generator plays your song back, keep the prompts and outputs as evidence. Find out what your state's publicity law covers. And check whether any contract you've signed claims rights in your voice or likeness, not just your recordings.

Update, 23 Oct 2026

AI Weekly, summarising an Axios scoop from 19 Sep, reports that the filing caught the USPTO and the Copyright Office off guard. On 18 Sep Universal and Sony filed a second suit against Suno over 60,202 recordings. It includes a claim that Suno stream-ripped audio from YouTube in breach of anti-circumvention law, which is an acquisition claim the DOJ brief doesn't reach. Suno's answer, covered here, still asserts fair use. On 21 Oct Reservoir's complaint argued that Suno's licence deals show a working market for training licences (our report). Footnote 13 of the DOJ filing reads like a direct answer to that argument.

Sources

• Statement of Interest of the United States, Doc. 316, filed 1 Sep 2026 (primary)
• US Copyright Office, Copyright and AI Part 3: Generative AI Training (May 2025) (primary)
• Landgericht München I, GEMA v. Suno press release (31 Jul 2026) (primary)
• Isbell et al. v. Suno complaint (31 Aug 2026) (primary)
• UMG Recordings v. Suno complaint (18 Sep 2026) (primary)
• Reservoir v. Suno complaint (21 Oct 2026) (primary)
• Bloomberg Law: Trump Administration Backs OpenAI in NY Times Copyright Suit
• NOTUS on the DOJ filing
• AI Weekly, summarising Axios (19 Sep 2026)