Computer scientists file amicus brief backing OpenAI: LLM training fits parameters, not copies of books

What happened

On 30 October 2026, a group of computer scientists filed a proposed amicus brief in support of OpenAI in In re OpenAI, Inc. Copyright Infringement Litigation, the multidistrict litigation in the U.S. District Court for the Southern District of New York (No. 25-md-3143, Judge Sidney H. Stein) over whether training large language models on copyrighted books and news articles is fair use. The brief (Document 2447-1) was filed on the court's 30 October amicus deadline, alongside briefs from publishers, music groups, authors' organisations and other interested parties.

What the brief argues

The computer scientists argue that LLM training is a computational use of text to fit numerical model parameters, not ordinarily a system for storing and retrieving the source works. A trained model captures statistical relationships across the corpus rather than holding addressable copies of the works, the brief says, and memorisation of training material — generating a near-exact replica — is limited, depends on how often text appears in the corpus, and is reduced by deduplication, post-training controls and input/output filters. The brief says the record shows OpenAI's safeguards are effective at preventing substantial extraction even under adversarial prompting. On that technical footing, it supports OpenAI's position that training is transformative.

Context

The 30 October amicus wave is heavily covered on this site, including briefs from the EFF, Wikimedia, the Authors Guild side, publishers and music groups, and the Suskind/Frye and Sag briefs. This computer-scientists brief is the distinctive pro-defendant technical voice in that wave and was not previously covered. It is a proposed brief filed with a motion for leave, not a court ruling, and it does not decide any issue in the case.