Fair use for AI training turns on facts, the Office says

The system

Whether training a music-generation model on copyrighted songs is lawful without a license is the question the recording industry keeps putting to courts, and courts keep answering slowly. The U.S. Copyright Office released a pre-publication version of Copyright and Artificial Intelligence, Part 3: Generative AI Training on 9 May 2025, ahead of a final version the Office says will not change its analysis. The report is the Office's own policy analysis of whether training generative models on copyrighted works, including music, qualifies as fair use; it is not a ruling on any pending lawsuit, and the Office says so directly. The Office's AI initiative page confirms the release date and situates Part 3 as the last of three parts opened by the same 2023 inquiry.

What the documents establish

The report declines to adopt a single rule for every AI training run. It states that courts must weigh the four fair-use factors together, with 'no mechanical computation or easy formula,' and that how much each factor adds depends on the facts of the particular case. It singles out the first factor, purpose and character of the use, including transformativeness, and the fourth, market effect, as likely to carry the most weight given the scale at which AI systems generate outputs. It also discusses licensing under the third factor: when licensing is reasonably available, unlicensed use of pirated or scraped works to compete in the market is less likely to be fair use, while non-commercial research that does not reproduce protected expression sits closer to fair. The Office is explicit: some training uses will be fair, some will not, and many uses will fall somewhere in between.

Craft and rights

For a rights holder negotiating a license with an AI music company, this report removes the comfort of a bright line in either direction. A developer cannot point to a blanket exemption for training, and a publisher cannot point to a blanket infringement finding; both arguments now rest on specifics the report lists, such as how training data was acquired, whether outputs can reproduce protected expression, and whether a licensing market already exists. This is an editorial read: the practical effect pushes disputes toward negotiated licensing or courtroom fact-finding, rather than settling the question once for the whole industry.

Outcomes and open questions

Because the analysis is fact-specific, the outcome for any AI music company depends on litigation the report does not decide, and on licensing markets it only describes as of its writing. Watch how courts weigh the first and fourth factors in music-specific cases now in discovery, and whether the final version of Part 3 changes anything beyond formatting.

• Did the training process reproduce your recognizable expression in outputs, or only learn general patterns?
• Was a license reasonably available for the training use in question, and was it sought?
• Which fair-use factor would a court likely weigh most heavily against your specific catalogue?

A fact-specific standard is not an answer for any single dispute; it is an instruction to keep the receipts on how a model was built and what its outputs can reproduce.