THE CRUNCH
AI coding assistants can churn out functional code at impressive speed, but a new study suggests the gains get stuck further down the pipeline. Harvard researchers Fiona Chen and James Stratton analysed aggregated analytics from Jellyfish, which tracks engineering team output, and found human code review acts as a significant bottleneck for firms using AI coding tools.
The dataset is substantial: 300 million individual work events such as commits and pull requests, plus issue management data covering more than 700,000 employees at over 700 software development firms, spanning 2021 through March 2026.
The study's authors report little evidence that firms using these tools increase software output or reduce employment. Any efficiency gained during coding, they find, is absorbed by downstream constraints: the review process grows significantly longer, pull requests are more likely to need revisions, and reviewers leave more comments. Experienced coders already know not to trust AI-generated code blindly, and the data suggests that caution has a measurable cost.