GCC bars most LLM-generated patches over copyright risk

The GCC Steering Committee has adopted a formal policy on artificial intelligence contributions that largely rejects patches produced by large language models, grounding the ban in unresolved copyright questions around training data and authorship.

David Edelsohn announced the decision for the committee, crediting a working group led by Jonathan Wakely. The text is now published on the GCC website. The project says the rules will be revisited periodically as the legal and practical picture changes.

Under the policy, legally significant LLM-generated changes are not accepted. Community discussion quickly fixed on a rough figure of about fifteen lines as the boundary of insignificance, though that number is not a hard limit written into the document. Separately, every contribution must be submitted by a human who understands the change and can answer questions about it.

The outcome surprised some observers who had read earlier project signals as more open to generative tools. Richard Biener said a conservative copyright stance was expected, but argued the policy "side-steps all important issues with LLM contributions by deflecting to legal issues," leaving aside reviewer overload, long-term maintainability, and grey areas for maintainers who only use models to prototype.

Wakely replied that contributions the policy already forbids need not burn reviewer time, and that any future relaxation would have to show it would not swamp maintainers or yield code nobody understands. Joseph Myers pointed to the human-understanding rule as the intended check on review burden. Working-group member Sam James said publishing a baseline was necessary to end a vacuum that had already complicated cases such as Google Summer of Code, even though stronger language lacked consensus and practical experience is expected to drive later revisions.