A recent empirical study from RIT analyzed 25,264 AI-generated pull requests across 2,361 popular GitHub repositories. The results challenge the hype around AI coding agents. While adoption is broad, depth is shallow: the median project received just 1-2 PRs over three months. Only 1% of projects saw AI contributions that exceeded human baseline productivity. Perhaps most telling, 78.9% of these PRs were reviewed, modified, and merged by the same developer who submitted them. This suggests that current 'human-AI collaboration' is largely a solo developer working with an agent, not a team of humans and AIs. For developer tool makers and AI engineers, this signals a gap: agents are easy to deploy but hard to integrate into real collaborative workflows. The study provides a sobering reality check for the AI coding narrative.
A study of 25,264 AI-generated PRs across 2,361 popular GitHub repos reveals that while AI coding agents are widely used, their impact is minimal. Median projects receive only 1-2 PRs over 3 months, and only 1% of projects exceed human baseline productivity. The finding that 78.9% of PRs are reviewed and merged by the same developer suggests current 'human-AI collaboration' is largely a solo developer plus agent model.