The Copyright Office's AI training report is still a draft 17 months on, and its critics are using that

On 9 May 2025 the US Copyright Office published the third part of its report on copyright and AI, the one about training, as a "pre-publication version". A note on its second page said the Office was releasing it early in response to congressional inquiries and stakeholder interest, and that a final version would follow "in the near future" with no substantive changes expected. Seventeen months later, there is still no final version. The Office's AI page lists Parts 1 and 2 as published, in July 2024 and January 2025, and says only that a final Part 3 "will be published in the future".

That might sound like a filing detail. It isn't. Part 3 is one of the most detailed official US statements in support of the argument music most needs to win, and the fact that it's unfinished has become a reason to ignore it.

What does the report actually say?

The report, 108 numbered pages, doesn't say training is or isn't fair use. It says it depends, and then says what it depends on:

• A spectrum. The Office expects some training uses to qualify as fair use and some not to (p. 74). Noncommercial research or analysis that doesn't let parts of the works show up in outputs is likely fair. Copying expressive works from pirate sources to generate unrestricted content that competes in the market, when licensing is reasonably available, is unlikely to be. It says many uses will fall somewhere in between.
• Sound recordings by name. On transformativeness, it says that using images or sound recordings to train a model that generates similar expressive outputs is hard to see as transformative, unless the original is being commented on or parodied (p. 46).
• Market dilution counts. The Office says the fourth fair-use factor, effect on the market, "should not be read so narrowly" (p. 65). The speed and scale of AI output risk diluting markets for works of the same kind as the training data. Its example of a diluted royalty pool comes from a Universal Music comment about cheap, generic AI content diluting human creators' royalties.
• Licensing markets matter. Where licensing can meet AI training needs, unlicensed uses "will be disfavored" under the fourth factor (p. 71). The report notes commenters' view that music is a sector well placed for expanded voluntary licensing (p. 68).
• Piracy weighs against. Knowingly training on a pirated or illegally accessed dataset should weigh against fair use, but doesn't decide it alone (p. 52).
• No new law yet. The conclusion calls government intervention "premature" and says licensing markets should keep developing, with extended collective licensing considered where gaps won't close on their own (p. 107).

Why is it still a draft?

The Register of Copyrights, Shira Perlmutter, was fired the next day. Publishers Weekly reported that the administration fired her on Saturday 10 May, two days after Librarian of Congress Carla Hayden was dismissed. Rep. Joe Morelle and others speculated that the report was the reason. No official reason was given, so that remains speculation.

Perlmutter sued. A federal appeals court let her stay in office while she challenged the firing, and on 30 June 2026 the Supreme Court denied the administration's request to remove her while the case continues. As Roll Call reported, the order says it is not a ruling on the merits. The Court had waited for its decisions in Trump v. Slaughter and Trump v. Cook, and IPWatchdog notes that Slaughter struck down the FTC's for-cause removal protections. Lawyers at Lowenstein Sandler wrote after the order that her staying on could lead the Office to formally publish the report. It hasn't yet.

Who cites it, and who attacks it?

It's easy to assume courts are leaning on the report. We checked, and that's not quite right. Lawyers cite it. Judges, in the two biggest rulings so far, didn't.

• Litigants, immediately. Three days after release, the authors suing Meta filed it as supplemental authority, arguing the Copyright Office is entitled to at least Skidmore deference.
• Not the June 2025 rulings. Neither Judge Chhabria's Kadrey v. Meta order nor Judge Alsup's Bartz v. Anthropic order cites it, by our reading of both. Chhabria embraced a similar dilution theory without citing the report, and still ruled for Meta, because the authors offered no evidence of it.
• Music publishers. In March 2026, Concord and the other publishers suing Anthropic over song lyrics quoted the report's dilution passage, royalty pools included, in their motion for partial summary judgment, and paired it with Kadrey.
• The Justice Department. Its 1 Sep 2026 filing in the OpenAI cases says the Register "appeared to endorse" the dilution theory, notes that she is challenging her removal, calls her reasoning "threadbare" and says, citing Loper Bright, that it doesn't warrant deference (footnote 17). We covered that filing here.

What we think

The report's unfinished status is doing work for its opponents. A document labelled pre-publication, from an official the executive branch says it removed, is easy to dismiss in a footnote. The DOJ did exactly that, while the Concord publishers present it as a conclusion the Office has reached. That fight is about the report's standing as much as its reasoning, and it's one the Office could settle quickly by publishing the final text it said needed no substantive changes.

For musicians, the content still matters more than the label. The report names sound recordings, cites a label's warning about royalty dilution, and treats a flood of competent generic tracks as harm that counts. That matches how streaming pays, as the Concord publishers' motion describes it: a fixed royalty pool split pro rata, so every stream an AI track earns shrinks everyone else's share. A court still has to accept that argument with evidence behind it, which is exactly what Kadrey lacked. The report won't win anyone's case for them.

The part we'd push back on is the licensing conclusion. "Let voluntary licensing develop" sounds neutral, but in music it means deals between AI firms and the companies that own the most catalogue. That's already happening: Suno announced a partnership with Warner Music Group in November 2025 to build new models on licensed music, and a Munich court's ruling against Suno gave Germany's GEMA leverage of its own. Nothing in the report makes sure the writers and performers whose work fills those catalogues get paid, beyond a brief nod to extended collective licensing for gaps. If the Office does finalise Part 3, the most useful revision would be one it promised not to make: saying who in a licensing market should actually get paid.

Sources

• US Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training, pre-publication version (May 2025) (primary)
• US Copyright Office, Copyright and Artificial Intelligence (primary)
• Kadrey v. Meta, Plaintiffs' Statement of Supplemental Authority (12 May 2025) (primary)
• Kadrey v. Meta, Order on cross-motions for partial summary judgment (25 June 2025) (primary)
• Bartz v. Anthropic, Order on fair use (23 June 2025) (primary)
• Concord v. Anthropic, Plaintiffs' Motion for Partial Summary Judgment (23 Mar 2026) (primary)
• Statement of Interest of the United States, In re OpenAI Copyright Litigation, Doc. 316 (1 Sep 2026) (primary)
• Suno, A new chapter in music creation (25 Nov 2025) (primary)
• Publishers Weekly, Copyright Chief Fired Amid AI Debate (May 2025)
• Roll Call, Supreme Court says Library of Congress official can stay in job (30 June 2026)
• IPWatchdog, SCOTUS Scraps Trump's Bid to Oust Perlmutter for Now (30 June 2026)
• Lowenstein Sandler via Legal 500, SCOTUS Ruling Re-opens The Door On The Copyright Office's AI Training Report (July 2026)