QEMU replaces blanket AI ban with pre-arrangement rules

QEMU is dropping its blanket ban on AI-generated contributions in favor of a cautious policy that lets maintainers gate large, heavily assisted work before it lands on the list.

Paolo Bonzini posted the fourth revision of the new rules on the QEMU development list, arguing the old prohibition no longer matches reality. Contributors have already sent AI-written patches and longer series, some of high quality, and even routine review can now produce agent-suggested fixes or test cases. A hardline reading made those everyday uses awkward while still leaving maintainers to police origin under the Developer Certificate of Origin.

The replacement is deliberate, not a free-for-all. Large or heavily AI-assisted contributions must be pre-arranged with the relevant subsystem maintainers and will face a higher bar. The goal, Bonzini wrote, is better code rather than faster dumps of dubious work, echoing Warner Losh’s warning against “free as in puppy” submissions that load present and future maintainers with technical debt.

Risks around provenance and originality are handled with layered controls: maintainer discretion and pre-arrangement, agent instructions that push the human to drive design and understand the output, and a new “AI-used-for” trailer that scopes where a Signed-off-by cannot fully guarantee control of origin. The accompanying agent guidance stresses a golden rule that the human should know the alternatives before implementation and should ask instead of letting the model choose.

The approach follows cautious experimentation elsewhere, drawing mainly on the Rust project’s policy and Software Freedom Conservancy guidance on LLMs, and keeps the focus on contributor-maintainer interaction rather than a pure legal firewall. Other projects have landed anywhere from social bans to narrow allowances for small patches and tests; QEMU is choosing the middle path while reserving the right for maintainers to set their own thresholds.