Why reviewing AI-generated code is devilishly hard(spinellis.gr)
spinellis.gr
Why reviewing AI-generated code is devilishly hard
https://www.spinellis.gr/blog/20260523/
3 comments
I agree with almost all of the article.
Reviewing the code yourself becomes almost impossible when we talk about hundreds of thousands of lines. In that case it might be easier to write the code yourself. What we can do is control and review the results.
I'm an electrical engineer working full-time as a project manager. I use AI after work, on weekends, vacations, or downtime to build something good, or apps I have in mind. If I had to learn programming properly, it would probably take me a few years, especially when you have kids and live far away from your family. So for me AI is my assistant, my coworker, my dev. I build what I have in mind and I control the results, not the code itself.
Example from today: I ran a database command and the AI reported it as done. But a row-level security rule silently blocked the write -> nothing was saved. It looked finished, but in reality it wasn't.
That's exactly the trap the article describes: the result looks plausible, so you stop checking.
Given that AI-generated code is not 100% perfect, we need to take into account the time and cost of reviewing when we discuss the productivity boost AI developers receive from code.