AI Code Generation Does Not Lead to Productivity Gains—Harvard Research

Even when using AI-powered code generation tools, the productivity of software development teams does not improve as expected—this is the finding announced by a research team at Harvard University. It has become clear that in development practice, the "amount of code written by AI" and "the amount of actual work completed" are not proportional.

This research was conducted on 718 companies based on data collected through the software development analytics platform "Jellyfish." Multiple companies of varying scales and industries were studied, making this a broadly representative survey of current development practice where AI utilization is advancing.

The survey revealed that when AI coding assistance tools or AI agents (autonomous AI systems that perform tasks) are introduced, the total volume of code generated does indeed increase. However, this did not translate into an increase in the overall completion of development work. The primary bottleneck cited is in the latter half of the development process, specifically the increased burden of "code review."

Code review is the process of human engineers examining and validating written code—an essential task to ensure quality. The more code AI generates, the more there is to review. As a result, subsequent stages become congested and overall progress stalls. In other words, it resembles a situation where widening only the entrance while keeping the exit at the same width does not speed up the flow.

This finding is significant in that it highlights "hidden costs" associated with AI adoption. While AI tools can automate and accelerate certain stages, the possibility that they introduce new burdens on downstream processes has not been adequately discussed until now. There is a risk that focusing solely on "code volume" as a productivity metric will lead to misinterpretation of actual conditions in the field.

When evaluating software development productivity, it becomes essential to take a comprehensive view of the entire process including not just the speed at which code is written, but also subsequent stages such as review, testing, and deployment. This research serves as one piece of evidence supporting this perspective.

As the adoption of AI coding tools accelerates, how companies will organize their review systems and redesign the entire development flow becomes a practical challenge for the field going forward. In tool selection and impact assessment, evaluation frameworks that focus on the quality and quantity of completed deliverables rather than on code generation volume are likely to become more aligned with reality.