Stanford study links code structure to AI productivity gains
A Stanford study of 120,000 developers found average AI productivity gains of 10%, but gains reached 300% in well-structured codebases. The findings suggest AI tools amplify existing code quality rather than uniformly improving output.
A Stanford study, described as being over a year old, observed 120,000 developers to measure productivity gains from AI usage. The average productivity gain across all developers was 10%, which is lower than individual experiences often suggest. The study identified strong correlations between software structure and AI effectiveness: teams working on clean, well-structured codebases achieved productivity gains of 200% to 300%, while teams with poor structure experienced slower development and increased production failures. The findings indicate that AI tools do not uniformly improve productivity but rather amplify the quality of the underlying codebase. Agents mimic existing conventions and patterns, meaning they will propagate architectural chaos if the codebase is already disorganized. The data suggests that AI adoption yields variable results dependent on pre-existing code quality, with structured codebases seeing significantly higher returns than unstructured ones.