AI coding now powers 30% of US code, but productivity gains favor experienced developers

AI coding work is shifting fast, and your career path may split

A new Science-backed analysis finds AI-assisted code is accelerating, but the biggest gains land with seasoned developers, not newcomers, changing how you should skill up. AI coding work is shifting fast, and the upside isn't landing evenly. A study published in Science suggests AI-assisted coding is now woven into everyday software creation on GitHub. The authors tracked Python development over time and used a purpose-built detector to flag code that likely came from generative assistants such as ChatGPT or GitHub Copilot. Recommended Videos For early-career developers, the signal is messy. Newer programmers lean on these tools more, yet the clearest performance gains show up among veterans. In other words, AI tends to reward people who already know how to steer it. How the study tracked AI code Instead of relying on surveys, the analysis focused on Python functions posted to GitHub and followed how individual developers changed over time. The method centered on a trained model designed to identify patterns associated with AI-generated code. That allowed the researchers to compare adoption across countries and experience levels, then connect usage to outcomes like commit activity and the range of libraries developers used. It watches what shows up in real repos, not what people say they do. Why experience changes the payoff Here's the hard part. Generative coding tools don't behave like a universal boost. Less experienced developers appear to use them more often, but the measurable gains concentrate among senior developers, including higher output and broader library usage. One likely explanation is judgment. Experienced developers tend to ask sharper questions, spot mistakes faster, and know when to ignore a plausible-looking answer. In that framing, AI coding work amplifies strong decision-making, not just speed. What you should do next If you're early-career, treat a copilot like a calculator, not a shortcut. Use it to draft boilerplate, explore unfamiliar libraries, and spin up tests, then make yourself explain every line you keep. Do that consistently and you'll learn faster than prompting alone. Check out the best AI tools for coding. If you're job hunting, build proof you can evaluate code, not just generate it. Clear READMEs, disciplined commits, and thoughtful code reviews will matter more as AI-assisted coding becomes normal. Watch for interviews that emphasize debugging and verification over memorized syntax.

AI used in one-third of newly written code in US by 2024: Report

AI is now a significant tool in software development, with nearly a third of new code in the US supported by AI systems by late 2024. While AI usage is high among less experienced developers, experienced programmers are reaping the productivity benefits. This trend could widen the gap between skilled and novice coders. By the end of 2024, around one-third of newly written blocks of computer programs in the US took support from AI systems -- gains in productivity, however, went to seasoned programmers, according to an analysis of patterns seen on a widely used collaborative coding platform. AI-supported coding was found to be high also in other countries, including Germany and France, with India catching up fast. "We analysed more than 30 million Python (programming language) contributions from roughly 160,000 developers on GitHub, the world's largest collaborative programming platform," author Simone Daniotti, from Austria's Complexity Science Hub and Utrecht University in the Netherlands, said. Trained on vast amounts of text data, generative AI can respond to a user's requests in the natural language, including writing and debugging computer code. Findings published in the journal Science also show that while AI usage is highest among less experienced programmers, productivity gains go to seasoned developers -- generative AI can therefore widen existing gaps, instead of levelling the playing field. Researchers used a specially trained AI model to identify whether blocks of computer code were AI-generated, for instance via ChatGPT or GitHub Copilot. The team was able to track programming work across the globe in real time as GitHub records every step of coding -- additions, edits, improvements. Python is among the most widely used programming languages in the world. "In the US, AI-assisted coding jumped from around five per cent in 2022 to nearly 30 per cent in the last quarter of 2024," author Frank Neffke, who leads the transforming economies group at Complexity Science Hub, said. "While the share of AI-supported code is highest in the US at 29 per cent, Germany reaches 23 per cent and France 24 per cent, followed by India at 20 per cent, which has been catching up fast," he said. Neffke added that Russia (15 per cent) and China (12 per cent) still lagged behind at the end of the study. "Currently AI writes an estimated 29 per cent of Python functions in the US, a shrinking lead over other countries," the authors wrote. The study also showed that the use of generative AI increased programmers' productivity by 3.6 per cent by the end of 2024. "That may sound modest, but at the scale of the global software industry it represents a sizeable gain," Neffke said. Less experienced programmers were seen to use generative AI in 37 per cent of their code, compared to 27 per cent for experienced programmers. However, productivity gains are driven exclusively by experienced users, the researchers found. "Beginners hardly benefit at all," Daniotti said, adding that generative AI therefore does not automatically level the playing field -- it can widen existing gaps. Experienced software developers were also seen to experiment more with new libraries and unusual combinations of existing software tools. "This suggests that AI does not only accelerate routine tasks, but also speeds up learning, helping experienced programmers widen their capabilities and more easily venture into new domains of software development," author Johannes Wachs, a faculty member at Complexity Science Hub, said.