Johannes Wachs of Budapesti Corvinus Egyetem and ELTE KRTK KTI reports that artificial intelligence (AI) is already contributing materially to software development. His Defacto analysis finds that nearly one third of program code is produced with AI assistance.
Major technology firms such as Amazon and Microsoft report that about 30 percent of code written within their ecosystems is created with AI tools. Wachs and colleagues set out to determine whether this pattern extends beyond Silicon Valley.
Method: open‑source code and machine learning detection
The study examined thousands of developers' open‑source work worldwide. The researchers trained a machine learning classifier to detect AI‑generated code in over 30 million Python files on GitHub. Between 2019 and 2024 they tracked 170,000 developers across six countries, measuring not only access to AI tools but actual usage frequency in everyday programming.
Adoption and cross‑country differences
Their results show rapid diffusion of AI tools and figures broadly consistent with corporate reports. By the end of 2024, they estimate that 29 percent of code in the United States was produced with meaningful AI assistance.
Key milestones correspond to noticeable jumps in AI‑written code: the introduction of GitHub Copilot in 2021, the appearance of ChatGPT at the end of 2022, and the 2023 wave of second‑generation models all coincide with clear accelerations. The US initially led adoption, but the gap has narrowed: Germany and France now sit around 23–24 percent AI share, India is catching up rapidly at about 20 percent, while Russia and China lag further behind—likely reflecting both supply constraints (limited availability of AI services) and demand‑side barriers such as state regulation or censorship.
Productivity effects: gains concentrated among experienced developers
To test whether AI makes programmers more productive, the analysis focused on developers in the United States and compared each developer to themselves before and after they began using AI tools, filtering out sector‑wide trends.
They find that increasing a developer's AI usage from zero to the current US average of 29 percent correlates with a 3.6 percent increase in the amount of code produced. While that average effect is modest, it becomes economically meaningful at scale. The productivity gains are uneven: experienced programmers show an average productivity increase of 6.2 percent, whereas less experienced developers do not exhibit a meaningful improvement. This suggests AI may widen rather than narrow performance differences among developers.
Changes in work patterns: more experimentation
Beyond quantity, the nature of work shifts. Developers who use generative AI more intensively tend to experiment with new software libraries and novel combinations more often, indicating greater willingness to explore unfamiliar technical areas. Most of these exploratory effects are also concentrated among experienced developers.
Economic estimate
The authors estimate annual US spending on programming‑related labor at $600–1100 billion. Applying their productivity estimates to current AI usage levels implies that AI is already generating $23–38 billion per year in value in US programming work alone (reported in the article as roughly 7,500–13,000 billion forints).
Conclusion
Defacto’s analysis concludes that change is well underway: AI is reshaping programming work. The pressing questions are not whether AI will transform the field, but how fast that transformation will proceed and who will be the winners and losers in the AI‑augmented future of software development.
