Post

Study Finds AI-generated Code Fails to Boost Productivity Due to Downstream Bottlenecks

A Harvard study of 718 firms finds that AI agents increased coding output much more than completed work: agents were associated with 30% more lines of code but only 12% more resolved issues. The authors point to downstream bottlenecks, especially review: average pull-request review time rose 49%, and the share of PRs needing changes doubled.

The study draws on about 300 million work events across more than 700,000 workers from 2021–2026, using Jellyfish platform data. InfoQ has no reader-comment section; its article relays Reddit reactions that more generated code creates review and maintenance liabilities, and suggests recovery time may be a more useful productivity measure than code volume. These are observational findings, not a universal causal estimate for every team.