Code review is burning out your best engineers
The New Stack Ankit Jain ● Covered by 7 sources
Best engineers are swamped reviewing AI code instead of writing it. The weird part: the more teams use AI, the more review slows down and burns out the people they rely on most.
Based on reporting by The New Stack, Ankit Jain — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Teams are finding that AI coding tools haven’t removed review work so much as moved it onto their best people. Senior engineers, especially the ones who care most about code quality, are getting buried in queues of machine-written diffs they do not enjoy checking. Some are even refusing to review AI-generated code at all.
The numbers in the source tell the same story. One study found that 77% of engineers now spend less time writing code, because more of their day goes to reviewing AI output. Teams with high AI adoption are merging 98% more pull requests, but review times have risen 91%. In some teams, the burden shows up as 15 pull requests a day, each around 400 lines long, landing on the same senior engineers.
What makes this harder is that AI code usually looks fine at first glance. It can be plausible but wrong, over-engineered, blind to a repo’s naming and logging conventions, confidently hallucinated, or built from cargo-cult patterns that mimic structure without understanding why. A human diff comes with context. The person who wrote it can explain the tradeoffs and constraints. AI leaves the reviewer doing archaeology on intent.
The source argues that the answer is not more heroic reviewing and not another LLM sitting on top of the first one. That just creates a bot-sitting exercise. Instead, it says teams should codify repeated feedback, preserve the reasoning trail from prompts and agent sessions, and measure the work that actually prevents bad code from slipping through.
One useful framing from the piece is that most review comments are not really judgment calls. The source says teams often find about 45% are deterministic, 30% are execution-testable, and 25% are genuine judgment, which means a lot of recurring comments could be turned into rules or checks. That leaves senior engineers doing the part that still matters: judging intent, scope, and blast radius instead of reading yet another 400-line cleanup job at 4 p.m.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI adoption people skip past when they brag about throughput. If the best engineers turn into permanent diff janitors, the whole setup is just a fancy way to burn out the adults in the room. Open or closed model, doesn’t matter much if the workflow still treats review as free labour and intent as disposable.
Read more about this at: The New Stack