One of the hardest things to do is review someone else's code. You can't know what they were thinking when they wrote it. It's also a challenge to find non-triv

One of the hardest things to do is review someone else's code. You can't know what they were thinking when they wrote it. It's also a challenge to find non-trivial bugs unless you spend almost as much time as the author getting a mental model of the code. Reviewing an AI's code strikes me as a pretty easy way to end up with a bunch of Heisenbugs and a team of developers that don't fully understand their codebase. If the internal expectation is a "25%" increase in development velocity a lot of engineers will just accept PRs with LGTM. If they heavily reviewed the code it might be perfectly fine but their number won't reflect the expected improvements and their quarterly or yearly reviews will suffer. You also can't go query that AI about why they wrote the code the way they did, not only is there inherent variability in response to prompts but an updated model might have entirely different responses with no way back to the original "author".