A Harvard University study analyzing 300 million work events across 700,000+ employees found that AI coding agents generate significantly more code but do not increase overall software output, as the efficiency gains are absorbed by bottlenecks in human code review. Reviewers leave more comments, pull requests require more revisions, and there is little evidence of increased output or reduced employment from AI tool adoption.
Background
This study by Harvard researchers Fiona Chen and James Stratton leveraged aggregated analytics from Jellyfish, tracking AI tool adoption across hundreds of firms from 2021 through March 2026, providing one of the largest empirical looks at real-world AI coding assistant impact.
- Source
- Ars Technica
- Published
- Oct 10, 2026 at 03:43 AM
- Score
- 7.0 / 10