Decisions · Policy · United States

Copyright and AI, Part 3: Generative AI Training

U.S. Copyright Office · 2025-05-09 · Proposal · non-binding

Pre-publication report, 9 May 2025

Can AI be trained on protected works?

Allowed with limits

Training is often transformative, but fair use is case by case and market dilution counts

Why hereAllowed with limits: the Office treats fair use for training as case by case, often transformative but weighed against market harm, including dilution.

The model as a copy

Is a trained model a copy of its training works?

A copy where it memorises works

Where a model has memorised works, copying its weights implicates the reproduction right

Why hereIn the middle: the Office says that where a model can output near-identical copies of works, there is a strong argument that copying its weights implicates the reproduction right for the memorised examples, and that the expression must exist in some form in the weights. It ties this to memorisation, not to every training work.

What it says

The Office finds some aspects of generative AI development likely amount to prima facie infringement, treats fair use as case-specific, reads market harm broadly (including dilution of the market for works of the same kind), and does not recommend statutory licensing for now.

Not binding on courts, but persuasive: US training is neither clearly allowed nor clearly banned.

Sources

Primary source ↗Summary ↗

This record

Data as of 2026-10-21.

Drafted with AI assistance and checked against the sources linked above; not yet reviewed by a lawyer. Methodology

Changes

2026-10-20Also coded on the new question on the model as a copy.