AI Contributions to GHC

GHC’s policy permits contributions made with large language models while placing responsibility on a human contributor. It prioritizes human communication with reviewers, requires merge requests to be clear, well explained, tested, and manageable in size, and says contributors must understand and be able to explain everything submitted. Directly generated code, documentation, tests, or other review material must be disclosed in the merge request or issue and commit message, typically with an “Assisted-By” line. Private use for research, debugging, or design exploration need not be disclosed. Writing the final version oneself is strongly preferred, especially for GHC Notes that preserve architectural reasoning, but it is not a requirement. Drafts and tickets receive more flexibility.

The policy’s rationale is that reviewer time, not the ability to produce code, is GHC’s scarce resource: a quick generated patch can impose hours of investigation, and plausible changes may conflict with years of project history. Contributions also help develop future maintainers, but GHC currently sees no large wave of low-quality AI submissions displacing human contributors, so a blanket ban would address a problem it does not yet face. Other projects choose stricter or more permissive rules, often reflecting legal concerns or community priorities; GHC’s approach treats disclosure as useful context rather than grounds for rejection. The policy therefore avoids judging LLMs in general and instead defines when assisted work is ready to consume review and maintenance effort.