When AI writes the code, how much of a developer's day goes to checking, correcting and making sense of it?

This explores how much of a developer's time goes to reviewing and fixing code that AI wrote. First, a straight answer: none of the notes here measures review-and-fix time as a share of the workday. What they do show is more useful. AI rarely removes work. It moves the work around, and the moved work can cost more than the work it replaced.

The clearest evidence comes from a randomized trial with experienced open-source developers working on their own mature codebases. With early-2025 AI tools, tasks took 19% longer, even though the developers had predicted a 24% speedup Do AI coding tools actually speed up experienced developers?. Low AI reliability was one of the named causes. When the output often needs checking or correcting, a developer who already knows the code well can lose more time verifying the AI than they would have spent writing it. The gap between predicted and actual results matters just as much. Developers and outside experts both expected gains, so the time spent on review is easy to underestimate, even while you're the one doing it.

Another result doesn't fit neatly with that one. A randomized trial with 96 Google engineers found AI features cut time on a complex task by about 21%, though the confidence interval was wide Do AI coding features actually speed up engineer productivity?. The two studies differ in setting: a single defined task at a large company versus developers working in codebases they knew deeply. That difference hints that the review burden depends on how much you already know. The more you know, the more the AI's mistakes stand out and the less it has to teach you. A broader note explains why both results can be true. AI tends not to reduce total task time. Instead it shifts time from doing the work to writing prompts and understanding what came back, which makes 'time on task' a weak measure of productivity Does AI really save time, or just change how we spend it?. In that sense, review and comprehension are where the reassigned time ends up.

What that review looks like varies with the person. In a study of students who were 'vibe coding' (building software mostly by prompting), 63.6% of interactions involved testing the running prototype and only 7.4% touched the code. Of those code interactions, 90% were reading rather than editing Where do vibe coding students actually spend their debugging time?. So some people don't fix AI code at all. They check whether the result works and go back to the AI when it doesn't. That leads to the least obvious finding. In a trial of developers learning a new library, AI use weakened their debugging ability and conceptual understanding. The exception was people who actively worked to understand the AI's output, who scored 65–86% on a follow-up quiz versus 24–39% for passive users Does AI assistance actually harm the way developers learn?.

The takeaway you might not have expected: how many minutes review takes is not the most important part. Skipping careful review makes you worse at reviewing and fixing code later. The time you save today by trusting the output may come back later, when you need to debug something the AI got wrong and the skills to do it have faded.