A class action over a generative model's training data makes provenance a litigation question

2022-11-03Copyright

Doe 1 v. GitHub, Inc., No. 4:22-cv-06823 (N.D. Cal., class action complaint filed Nov. 3, 2022), against GitHub, Inc., Microsoft Corporation, OpenAI, Inc., OpenAI, L.P., OpenAI GP, L.L.C. and three OpenAI Startup Fund entities

Two anonymous programmers filed a class action complaint in the United States District Court for the Northern District of California against GitHub, Inc., Microsoft Corporation, OpenAI, Inc. and five affiliated OpenAI entities over GitHub Copilot and the OpenAI Codex model behind it. The complaint alleges that the defendants "have conceded that the training data includes data in vast numbers of publicly accessible repositories on GitHub," that those repositories carry open-source licenses whose common term requires attribution, a copy of the license and the author's copyright notice, and that Copilot "often simply reproduces code that can be traced back to open-source repositories" while "code reproduced by Copilot never includes attributions to the underlying authors." The twelve counts are violation of the Digital Millennium Copyright Act, Sections 1201 through 1205 of Title 17 of the United States Code, for removal of copyright management information; breach of the open-source licenses; tortious interference; fraud; false designation of origin under Section 1125 of Title 15; unjust enrichment; unfair competition; breach of GitHub's privacy policy and terms of service; violation of the California Consumer Privacy Act; negligent handling of personal data; civil conspiracy; and declaratory relief. No count of copyright infringement is pleaded. The complaint alleges the defendants have offered "shifting accounts of the source and amount of the code or other data used to train and operate Copilot" and have cited fair use.

What it changesThe theory does not depend on whether training is fair use; it depends on what the model reproduces and on what the defendants can say about the data that went in, and the complaint already turns the defendants' own statements about training data against them. For a company that trains or fine-tunes its own models, the in-house action is a provenance record: for each dataset, where it came from, under what license or terms, what was filtered out, and who approved it, kept as an engineering artifact rather than a legal memo so that it exists before anyone asks. The same record is what a patent application on the training method needs for its written description, since a claim to training on a corpus selected or filtered in a particular way must describe the selection. Engineering teams should also treat verbatim reproduction of training data as a measured defect, because that measurement is the evidence in a case like this one.

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.