US DOJ tells the NYT v OpenAI court that training LLMs on copyrighted works is fair use
On 1 September 2026 the United States filed a Statement of Interest (ECF 1682) in In re OpenAI, Inc. Copyright Infringement Litigation, MDL 25-md-3143 (SDNY, Judge Stein), arguing that copying written works to train a large language model is 'exceedingly transformative' fair use, that the plaintiffs' 'market dilution' theory of harm is 'deeply flawed', and that licensing barriers would hand the largest companies an oligopoly on LLM training. The brief binds no one; the court can weigh it as it sees fit. What it does is put the executive branch formally on the AI providers' side across every publisher and author claim in the MDL, ahead of summary judgment. For any organisation licensing or building on US-trained models, the practical reading is that compensation for training data will be won through licensing deals or legislation, not from a judgment. Note the contrast with Europe, where text-and-data-mining sits on an opt-out right under the DSM Directive and the AI Act's GPAI copyright duties, not on a fair-use defence.
Unsure which of these developments applies to your AI systems?
Begin in writing →This briefing is general information, not legal advice, and does not create an advisor–client relationship. Summaries are original; follow source links for the full record. Adesanya AI Advisory — Abdulwahab B. Adesanya, Barrister-at-Law (Nigeria).