AI boosts code output, but human reviews slow delivery

A Harvard University research paper found that using AI increased the amount of code developers produced, but did not lead to a statistically significant improvement in completed software. The paper’s authors, Fiona Chen and James Stratton, reported a 30% rise in lines of code, a 20% rise in code commits and a 23% rise in pull requests.

Review work also increased. Time to review and merge pull requests rose by 49%, reviewer comments per pull request increased by 35%, and the share of employees doing code reviews grew by 14%. The average review process lengthened from 7.03 days to 10.48 days.

The researchers suggest that AI-generated code requires more corrections, shifting the bottleneck to code review rather than removing it. They also raise the question of whether companies are assigning more workers to reviews as AI changes workloads. The paper says humans remain heavily involved in reviewing code, even as AI is used in that stage.

The report frames the challenge as finding ways to use AI to support employees and reduce unnecessary workload. It does not report significant job losses from AI adoption. Its conclusion is that productivity gains are absorbed by downstream constraints, with code review becoming the bottleneck.