The US Copyright Office says licensing for AI training should be left to grow
Its report on generative AI training finds licensing deals spreading, and counts lost licence income as real harm to owners.
Picture made for ABURA Bureau.
On 9 May 2025 the US Copyright Office released a pre-publication version of its report on generative AI training, the third part of its study of copyright and artificial intelligence.
On harm to owners, the report counts lost licensing income as part of the picture, and says a licensing market "need not be long-standing or exhaustive" to count. It records that by the end of 2023 AI developers were already licensing works in sectors including music, vocal recordings and news.
Its policy conclusion is plain. Licensing agreements for AI training, individual and collective, are "fast emerging in certain sectors", though so far unevenly, and the Office judged that government intervention would be premature. Where gaps are unlikely to close on their own, it points to options such as extended collective licensing.
The report is advice, not law, and courts decide fair use case by case. But it sets out the view that a market for training data exists and that bypassing it can hurt an owner, a view the Third Circuit took up in Thomson Reuters v. ROSS in September 2026.
