The first suit over training on other people's code
On 3 November 2022 two plaintiffs suing under pseudonyms filed a class action in the Northern District of California against GitHub, Microsoft and OpenAI, case 3:22-cv-06823. The complaint says Copilot, trained on public repositories, emits other people's code without the three things eleven widely used open-source licences require: attribution, a copyright notice, and the text of the licence itself.
Why it matters
Until then no United States court had been asked whether training a commercial model on work posted in the open is lawful. The suit moved the question from argument into proceedings, and it chose not copyright but section 1202 of the DMCA, on the removal of copyright management information, because an open licence permits copying and does not permit removing the author's name. Every suit of 2023 grew out of that construction.
The eleven licences are named one by one in the complaint: Apache 2.0, GPL-3.0, MIT, BSD 2, BSD 3, BSL-1.0, EPL-2.0, AGPL-3.0, GPL-2.0, LGPL-2.1 and MPL-2.0. The plaintiffs compute minimum statutory damages as 2,500 dollars per violation, three violations for each unlawful Copilot output, across the 1.2 million Copilot users Microsoft reported in June 2022: 3,600,000 violations and over 9 billion dollars. On 10 November the same lawyers filed a second complaint for two further plaintiffs. What the record does not claim. The complaint pleads DMCA section 1202 without naming a subsection; 1202(b) is not in it, and 1202(c) appears once as the definition of copyright management information. Nothing is said here about how the case ended.