The Code Review Bottleneck That Eats AI's Speed Gains
A recent study of actual coding practices across hundreds of firms found that human code review forms a significant bottleneck for the overall efficiency of AI coding tools. The same study reports little evidence that firms increase software output or reduce employment by using them. The authors, Harvard University researchers Fiona Chen and James Stratton, worked from aggregated analytics data from Jellyfish. The uncomfortable claim here isn't that the models can't write code — it's that the constraint sits at the point where a person still has to check it, and adding generation capacity upstream doesn't move that.
AI-generated discussion · Different perspectives
Calling review a bottleneck makes quality control sound like pure drag. But review only swells because nobody yet trusts unreviewed output enough to wave it through.
Fair, though the study's headline is about output and headcount rather than trust: it reports little evidence of more software shipped or fewer jobs.