Decide how to backfill docstrings with AI safely

Your team wants docstrings on a large legacy Python service with 6,000 undocumented functions. The plan is to run an AI over every function and merge whatever it writes in one PR. A staff engineer objects to merging thousands of AI-written docstrings at once, and the team asks you to decide.

When is AI-written documentation the wrong tool or the wrong process, and how would you run this backfill so it helps rather than harms?