On 9 May 2025 the US Copyright Office published the pre-publication version of part three of its Copyright and Artificial Intelligence report, on generative AI training. Its subject is narrow and pointed: when does using copyrighted works to train a generative model amount to fair use?
The report offers no single rule; it frames training as a case-by-case question in which the commercial character of the use and its effect on the market for the original work are central. For companies, the practical step is to keep records of corpus provenance and licences, so that the character of the use can be shown if it is later challenged.
Why the report matters
The Copyright Office is not a legislature, and the report is not directly binding, but it supplies an analytical framework for the litigation that has followed. Where courts, regulators and companies are each saying something different, a systematic analysis from the responsible authority significantly shapes where the discussion starts.
For AI companies targeting overseas markets, it bears directly on how training data is procured and retained: if the boundary of fair use is uncertain, risk has to be reduced through an evidence trail.
The report’s analytical logic
The report reaches neither a conclusion that training is always lawful nor one that it always infringes. Instead it breaks the question into variables: the purpose and character of the use, the nature of the original work, the amount used, and the effect on the potential market for the original.
The most critical is the market substitution judgement. Where model output can substitute for the original in the market, the room for fair use narrows. This echoes recent judicial trends in Europe. See our note on the GEMA v OpenAI ruling in Munich.
Separating the training question from the output question
Training and output are two independent questions. Even where the training step is lawful, output may still infringe by being substantially similar to an existing work. China already has rulings on copyright in AI-generated content. See our notes on the Shanghai AI prompt copyright case and on the Beijing court criminal copyright case involving AI.
Where companies can start
First, build a training data ledger, recording sources, licence status and permitted scope, so that evidence for a fair use defence exists.
Second, keep compliance judgements for training and for output separate, and do not apply the conclusion from one to the other.
Third, watch the market substitution risk of output, and set filters for functions that may reproduce a particular work’s style or content. See our notes on the patent examination guidelines on AI and on the Thaler Supreme Court case on AI authorship.
There is no once-and-for-all answer on the lawfulness of training data, but it can be demonstrated. Those who can explain where their data came from hold the initiative in litigation and due diligence.
This article is based on part three of the US Copyright Office’s Copyright and Artificial Intelligence report and public information and is for reference only; it does not constitute legal advice.