Study finds AI wrote about 30% of US developers' Python code by late 2024

Researchers Simone Daniotti, Johannes Wachs, Xiangnan Feng and Frank Neffke trained a classifier to detect AI-generated Python functions across more than 31 million GitHub commits from roughly 170,000 developers, then used it to track adoption over time and by country. They found that by the end of 2024, AI had written an estimated 29% of new Python functions committed by US-based developers, with Germany and France close behind at around 23–24%; adoption in China lagged well behind the leading countries.

The paper, later published in Science, estimated that AI-assisted coding had lifted quarterly output — measured by online code contributions — by roughly 3.6%, translating to tens of billions of dollars in additional value annually across the developers studied. The distribution of that benefit was uneven: adoption of AI coding tools was highest among less experienced programmers, but the productivity gains accrued disproportionately to seasoned developers. The authors concluded that generative AI did not level the playing field between novice and expert coders as some had predicted, and in some respects widened it, since experienced developers were better able to catch and correct AI-generated errors.

The study offered one of the more methodologically grounded estimates of AI’s real-world penetration into a specific knowledge-work task, based on directly observed code rather than self-reported survey data, and became a frequently cited data point in debates over AI’s effect on software engineering employment and skill development.