Scott Chacon posted an RFC patch following the AI discussion at the contributors' summit. He says the current text encourages careful use of AI but also says Git will reject anything that looks AI-generated. In his view that leaves contributors with useful, reviewed work unsure whether disclosing their tools will get the patch rejected. He wants to move closer to the Linux kernel's approach: use the tools you find helpful, but take responsibility for what you send.
The patch would allow AI-assisted contributions under the usual quality and licensing requirements. It would require human understanding, appropriate testing and disclosure of substantial assistance. The DCO would stay unchanged, contributors would have to consider provenance and meet license obligations, and reviewers could ask for more evidence or decline work they cannot confidently assess. The appearance-based rejection language would be replaced.
brian m. carlson said he does not favor the policy. His argument is that the major models were trained on code and text from many sources, including ones that forbid copying, modification or distribution, and that models are known to reproduce parts of their training data. He asked how anyone can honestly assert the DCO without knowing every source involved. He added that even MIT-licensed training data would still require a copyright and permission notice on every copy.
Chacon replied that generated output can reproduce material Git lacks permission to distribute, though he thinks that is very rare for anything complex. He asked whether the possibility means contributors must know every training source before certifying the DCO. He noted that human contributors have also read code under many licenses and are not asked to account for it, but are expected to have the right to submit their contribution and to respect the licenses of material they incorporate.
Junio C Hamano said Chacon's framing makes the proposal sound as legally sound as the kernel's, but the only assurance offered is that an OpenAI model produced plausible-sounding text, unless Chacon is willing to act as Git's lawyer. He said the proposed text tells developers and reviewers to do things quite different from the kernel's rules, and that little in it seems meaningfully enforceable by reviewers. He said he was partly playing devil's advocate.
Luca Milanesio distinguished learning from existing code from copy-pasting it, and argued that LLMs do the latter statistically. Kristoffer Haugsbakk objected to analogies between human learning and models, and said an individual who wants to avoid copyright risk is entitled to. No consensus has emerged, and the patch remains an RFC.