I think this take about rejecting AI-written PRs unless the developer understands every implementation detail is silly. Some of these arguments increasingly sound like developers trying to protect their relevance.
We already build on layers of abstraction we don’t fully understand. Follow that argument far enough and you shouldn’t submit a PR unless you can explain the bytecode it produces.
The focus should be on why the change is needed, what guarantees we have that it works, and what evidence we have that existing behaviour is preserved.
Understanding the implementation helps establish those things. But “I understand every line because I wrote it” has never been proof of correctness.