Harvard Paper Finds AI Coding Output Gains Are Absorbed by Code Review Bottleneck
AI lifts coding output 30%, but a Harvard paper finds no significant completed-software gain, as code review bottlenecks.
The paper, by Fiona Chen and James Stratton, measured higher output across several stages. Lines of code produced rose 30%, code commits rose 20%, and pull requests rose 23%, TechRadar reported. Yet the researchers found no statistically significant improvement in finished software, prompting them to examine where the productivity was being absorbed.
They found delays and workload shifted to review. The time taken to review and merge pull requests increased 49%, and the average review process lengthened from 7.03 days to 10.48 days. TechRadar's report gives conflicting figures for reviewer comments per pull request: one summary line says comments per pull request rose 49%, while the article body says the increase was 35%. The paper also found that the proportion of employees doing code reviews rose 14%, according to the report.
The researchers argue that these patterns show AI is moving the bottleneck further down the production process rather than removing it. Although AI is already used to some degree in code review, TechRadar reported that humans are still heavily involved in that work. The paper also questions whether companies are responding to the shift by assigning more workers to code review.
The report connects the findings to broader debates about AI and jobs. Layoffs have often been associated with AI adoption, whether justified or not, and concerns about job losses continue. However, TechRadar said the paper indicates human workers remain vital, just in different parts of the coding process. It also notes that AI adoption has not resulted in significant job losses, according to the report, making the practical question less about replacing workers outright than about how companies use AI to support employees and reduce unnecessary workload.
The paper also argues that AI-generated code needs substantially more corrections, raising the question of whether the technology is helping as much as raw output figures suggest. Its conclusion, as quoted by TechRadar, is that "The productivity gains are absorbed by downstream constraints in the production process," and that "code review becomes the bottleneck."