AI-first company

Study

Across 160,097 developers, experienced programmers gained from AI coding tools and early-career ones did not

AI wrote an estimated 29% of Python functions in the United States. Quarterly output, measured in code contributions, rose an estimated 3.6%. Experienced developers produced more and moved into new kinds of software; early-career developers showed no clear benefit.

By
Last reviewed
Type of study
Data analysis
Published
2026
Corrections
Report an error

At a glance

Source
Daniotti, Wachs, Feng and Neffke (2026). Who is using AI to code? Global diffusion and impact of generative AI. Science, 391(6787), 831–835. doi:10.1126/science.adz9311
How it was done
The authors trained a program to recognise Python code written by AI. They ran it over public code on GitHub to track how fast AI coding spread, and estimated how developers' output changed as they took it up. Data analysis. It builds a measure from existing data, such as job and task records.
Who took part
More than 30 million commits to GitHub by 160,097 software developers.
Where
Public Python code on GitHub, worldwide, with country comparisons.

“By contrast, early-career developers showed no significant benefits from AI adoption.”

What it found

  • How much AI writes. AI wrote an estimated 29% of Python functions in the United States. The US lead over other countries is shrinking.
  • Output. Quarterly output, measured in code contributions, rose an estimated 3.6% because of AI.
  • Who gained. Experienced, senior developers who used AI produced more. They also moved more readily into new areas of software.
  • Who did not. Early-career developers who used AI showed no clear benefit.

An earlier version of the paper on arXiv, from November 2025, gave the sample as about 170,000 developers (arXiv 2506.08945). The figures on this page come from the journal version.

When the result holds

The study follows developers’ public work over time, outside any one company. Studies inside companies point the other way. In customer support, for example, less experienced agents gained most from an AI assistant (Brynjolfsson, Li and Raymond, 2025). Those studies measure output on set tasks with the tool. This one measures what developers chose to build in the open. The evidence on who gains most points both ways.

Limits

  • The study observes developers; it does not assign AI by chance.
  • It relies on a program that recognises AI-written code, which can be wrong.
  • Output means code contributions, not the value of the code.
  • It covers public Python code only.

Practices that use this study

This study limits or contradicts: