Interpretation of learning effects from AI tool experience

Determine the causal impact of developers’ accumulated Cursor experience on observed speedup or slowdown and disentangle whether any apparent speedup at higher experience levels arises from improved AI-assisted performance or atrophy of non-AI performance when working on large open-source repositories.

Background

The authors analyze speedup by hours of Cursor experience and find little change up to 50 hours, with one developer beyond 50 hours showing positive speedup. They caution that this could reflect either genuine improvement with AI-assisted workflows or deterioration in non-AI performance due to skill atrophy.

Because the analysis is underpowered and confounded by potential atrophy effects, the true direction and magnitude of learning effects on productivity remain unresolved.