Abstract:Generative coding tools promise big productivity gains, but uneven uptake could widen skill and income gaps. We train a neural classifier to spot AI-generated Python functions in over 30 million GitHub commits by 170,000 developers, tracking how fast -- and where -- these tools take hold. Today, AI writes an estimated 29% of Python functions in the US, a modest and shrinking lead over other countries. We estimate that quarterly output, measured in online code contributions, has increased by 3.6% because of this. Our evidence suggests that programmers using AI may also more readily expand into new domains of software development. However, experienced programmers capture nearly all of these productivity and exploration gains, widening rather than closing the skill gap.
Submission history
From: Johannes Wachs [view email]
[v1]
Tue, 10 Jun 2025 16:06:19 UTC (1,552 KB)
[v2]
Thu, 20 Nov 2025 13:29:03 UTC (4,472 KB)