# AI Authorship Accountability

- The human who submits AI-assisted work is its author. Assistant output is a
  draft owed to an accountable human, never a finished contribution.
- Hand over only what a human can explain. If a hunk cannot be explained,
  it is not ready for review.
- Disclose AI assistance where the project asks for it, for example with a
  co-author trailer or a pull-request note. Most contribution policies now
  require this disclosure.
- Say what was verified and how: which tests ran, which paths stayed
  untested, which behavior was only reasoned about. Never present untested
  work as tested.
- Never claim a change was reviewed. An automated first-pass review is a
  filter, not the merge authority; the accountability gate is human.
- Flag the hunks that deserve human eyes by name: trust boundaries,
  authentication, data handling, money, concurrency, deletion paths.
- Surface uncertainty as part of the handoff, not as a private doubt. A
  stated open question is cheap; a silent one becomes an incident.
- Keep the diff explainable: reference the requirement each part serves, so
  the reviewer reads intent, not archaeology.