Study: AI Code Generation Does Not Increase Software Output
A new study reveals that while AI coding agents generate more code, they do not increase overall software production or reduce employment. Human code review becomes a bottleneck, consuming efficiency gains.
AI coding assistants and agents are highly efficient at generating large volumes of functional code, but this does not translate to increased software output or reduced employment, according to a recent study from Harvard University. The research found that human code review acts as a significant bottleneck, absorbing any efficiency gained during the coding phase.
The study, which analyzed aggregated engineering analytics data from over 700 software development firms between 2021 and March 2026, utilized data encompassing over 300 million work events such as commits and pull requests. The findings indicate that while the act of coding is accelerated by AI, the overall software production process does not see a proportional increase in efficiency.
Researchers Fiona Chen and James Stratton observed that the implementation of AI coding tools led to an increase in the length of code review processes. Pull requests were more likely to require revisions, and reviewers left more comments, effectively extending the time spent downstream from the initial code generation.
This suggests that the perceived gains from AI in speeding up initial code creation are offset by the increased demands on human oversight and quality assurance. The study highlights that firms using these tools have shown little evidence of increased software output or reduced employment, indicating that human review remains a critical, and now potentially slower, part of the development cycle.