> That is less useful when the changes are editing the tests but we don't know if a human has validated the assertions.

Yep. I've witnessed this first hand many times. AI-enthusiastic coworker submits a PR. The tests don't pass. "Can you fix the tests? Then I'll review."

Next commit has `assert status == 200` changed to `assert status == 500` all over the place, among other things. Yes, technically, the tests now pass, but...

Last summer, this went on with one guy for weeks. Thousands and thousands of lines of slop. Eventually he was moved off the project and we threw away all his changes.