The New York Times sues Microsoft and OpenAI and files pages of GPT-4 output reciting its articles verbatim

2023-12-27Copyright

The New York Times Company v. Microsoft Corporation, OpenAI, Inc., OpenAI LP, OpenAI GP, LLC, OpenAI, LLC, OpenAI OpCo LLC, OpenAI Global LLC, OAI Corporation, LLC, and OpenAI Holdings, LLC, No. 1:23-cv-11195 (S.D.N.Y., complaint filed Dec. 27, 2023) (assigned to Stein, J.)

The New York Times Company filed a 69-page complaint in the Southern District of New York against Microsoft Corporation and eight OpenAI entities, assigned to Judge Sidney H. Stein, pleading seven counts: direct copyright infringement under Section 501 of Title 17 of the United States Code against all defendants, vicarious infringement, contributory infringement against Microsoft and in the alternative against all defendants, removal of copyright management information under Section 1202, common law unfair competition by misappropriation, and trademark dilution. The complaint alleges that the GPT models were trained on unauthorized copies of Times works, that www.nytimes.com is the most highly represented proprietary source in a filtered English-language subset of a 2019 snapshot of Common Crawl, "the third overall behind only Wikipedia and a database of U.S. patent documents," and that the models have "memorized" copies of those works encoded in their parameters. Exhibit J collects scores of examples in which GPT-4 outputs near-verbatim copies of significant portions of Times articles, including a 2019 taxi medallion investigation. It also alleges that products built on Microsoft's Bing search index return verbatim excerpts and detailed summaries of Times articles and attribute false information to the Times. The prayer seeks statutory and compensatory damages, a permanent injunction, and "destruction under 17 U.S.C. § 503(b) of all GPT or other LLM models and training sets that incorporate Times Works," with a jury demand.

What it changesExhibit J is a demonstration that memorization is discoverable from the outside: a plaintiff with a few prompts can show what a model was trained on without discovery. For a company whose patent applications describe training on licensed, proprietary or synthetic data, that description is now a factual assertion an adversary can test against the model's outputs, and a specification that overstates the cleanliness of the training set is a credibility problem the first time an engineer is deposed. Treat training-data provenance as part of the patent file: record the sources, licenses and filtering for each model named in an application, and have the specification describe the pipeline in terms the record supports. The requested remedy, destruction of the models and training sets, is a reminder that the model weights are the asset at risk, which argues for claiming the architecture, training method and inference pipeline rather than any one trained instance.

More on Copyright and training data · Litigation, damages and PTAB challenges

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.