Experienced Developers Saw AI-Linked Gains in Output and Exploration

Software development is becoming one of the clearest places to watch a new kind of human capability take shape.

By the end of 2024, an estimated 29 percent of Python functions contributed by US developers to open-source projects showed substantial AI support. France and Germany were close behind at 24 and 23 percent, while India had reached 20 percent.

Those figures come from a peer-reviewed *Science* study that examined about 31 million contributions and five million Python functions from 160,097 developers across the United States, China, France, Germany, India, and Russia. The observation window ran from 2019 through 2024.

The researchers did not rely on surveys or direct logs from coding assistants. They trained a neural classifier on human-written functions and synthetic AI-generated equivalents, then used its predictions to estimate AI-assisted code in public GitHub activity. The classifier achieved an out-of-sample ROC AUC of 0.96, although its accuracy was somewhat lower on code from newer models and real-world model interactions.

Research references: [1]

More activity and a wider technical range

For 100,097 US developers, the researchers used fixed-effects models to compare each person’s activity across quarters as their estimated AI use changed. At the 29 percent adoption level observed at the end of 2024, greater estimated AI use was associated with a 3.6 percent increase in quarterly commit rates.

The association was larger among senior-level users. For that group, moving to the same 29 percent adoption level implied a 6.2 percent increase in commits. The paper reported no statistically significant productivity effect among early-career users.

Estimated AI use was also associated with broader experimentation. At the same adoption level, developers were predicted to use 2.7 percent more new combinations of software libraries. The result was robust to several alternative definitions of new library use, including analyses limited to common libraries and broader library categories.

Libraries are reusable building blocks for work such as visualization, databases, web interaction, and natural-language processing. Trying a new combination can signal movement into technical territory a developer has not used before. It is a useful proxy for exploration—not proof of innovation, better software, or business value.

Research references: [1]

Experience appears to change what people get from AI

The study’s most consequential pattern was not simply who used AI most.

Developers who had only recently joined GitHub had an estimated 37 percent AI-supported share, compared with 27 percent among the longest-tenured users. GitHub tenure is an imperfect proxy for professional experience, but on this measure the newer developers were heavier adopters.

Yet the measurable productivity and exploration associations were concentrated among senior-level users. Early-career estimates were not statistically distinguishable from zero. The paper argues that experienced developers may be better able to interpret generated code, identify mistakes, and connect new components to a larger system. Those mechanisms are plausible explanations, not effects the study tested directly.

The pattern complicates the idea that access alone levels the playing field. Powerful tools can make knowledge easier to reach, but they do not automatically supply the judgment needed to evaluate and integrate what they produce.

Research references: [1]

The career ladder still matters

If experience changes the value people obtain from AI, the challenge is not simply to give junior developers more AI. It is to help them build the judgment that experienced developers already bring.

Traditional apprenticeship develops that judgment through small tasks, code review, debugging, and exposure to other people’s reasoning. AI can perform some entry-level tasks quickly. Removing those tasks without replacing the learning inside them could weaken the route to mastery.

A better design would make reasoning visible. New developers could explain why they accepted or rejected a suggestion, test generated code against edge cases, compare implementations, and trace how a new library changes system behavior. Senior developers could use AI-assisted work to expose decisions rather than merely deliver finished answers.

The study did not evaluate such a training program. This is a design implication, not an experimental result. But it follows from the central human-agency question: access to a tool is not the same as developing the person who directs it.

What this evidence can—and cannot—establish

The analysis is unusually large, uses within-developer comparisons, and is accompanied by a public Dryad replication package containing the underlying analysis inputs, model outputs, and code-related data structures.

But it remains observational. Fixed effects control for stable differences between developers and shared quarterly patterns, yet other changes can still move alongside estimated AI use. The classifier measures statistical traces of AI assistance rather than direct provider telemetry, and measurement error may affect the estimated associations.

The data cover open-source Python work on GitHub, not every programming language, company, or country. Commit counts measure activity, not code quality. The study does not establish fewer defects, stronger testing, faster issue resolution, better security, or more valuable products.

The current Science record lists January 22, 2026 as the online publication date, February 19 as the issue date, and identifies this as the most recent article version. It reports Austrian and Hungarian public research funding, no competing interests, and a replication dataset on Dryad.

Research references: [3] [1] [2]

Human agency is the multiplier

The strongest supported conclusion is not that AI made every developer better. It is that people differed substantially in what they appeared able to do with the same emerging capability.

AI can offer syntax, alternatives, and unfamiliar building blocks at extraordinary speed. Human experience supplies context, standards, responsibility, and the ability to know when a plausible answer is not a good one. In this dataset, greater AI use tracked with more output and broader technical exploration for experienced developers.

The opportunity now is to make that partnership teachable. The future of useful work should not divide people into those whose expertise is amplified and those who merely consume generated output. It should create better routes for more people to develop the judgment that turns assistance into agency.

: Daniotti, S., Wachs, J., Feng, X., & Neffke, F. “Who is using AI to code? Global diffusion and impact of generative AI.” *Science* 391, eadz9311 (2026). https://real.mtak.hu/234712/1/science.adz9311.pdf : Science journal record for DOI 10.1126/science.adz9311, accessed October 21, 2026. https://doi.org/10.1126/science.adz9311 : Daniotti, S., Wachs, J., Neffke, F., & Feng, X. Replication data for “Who is using AI to code?” Dryad, version published January 21, 2026. https://doi.org/10.5061/dryad.3r2280gv0

Research references: [1] [2] [3]