Scientifically Speaking: AI is writing code and changing jobs. Who actually benefits?

The message was unmistakable. India is gearing up to be a major player in the AI era. At the same time, a study published in Science offers insight into how AI is actually being used to write software. Software runs everything from your banking app and hospital records to the railways booking system. The people who write and maintain it are among the workers most directly affected by AI. What is happening to them now may foreshadow what is coming for knowledge workers everywhere.

The study was led by Simone Daniotti at Harvard’s Growth Lab, with colleagues at the Vienna University of Economics and Business and the Complexity Science Hub. The researchers trained a machine learning classifier to identify AI-generated code in more than 30 million contributions written in Python, a widely used programming language, by 160,000 developers on GitHub, the world’s largest platform for sharing and managing software. They tracked adoption across six countries from 2019 to the end of 2024.

By late 2024, AI was generating roughly 29% of programming functions in the US in the GitHub repositories the researchers analysed. Germany and France were close behind at 23 to 24%.

India was catching up fast at 20%. When the researchers compared the same developer’s output before and after adopting AI tools, they found a 3.6% increase in quarterly productivity. That sounds modest until you consider the scale. Even conservatively, AI coding tools appear to be generating about $10 to $15 billion a year in additional value in the US alone, with higher-end scenarios far larger.

But the most surprising finding was about who benefits. The productivity gains were concentrated almost entirely among experienced, senior-level developers with six or more years of activity. They wrote more code and experimented more, branching into unfamiliar technical territory. Early career developers, despite using AI more frequently than their senior colleagues, showed no measurable productivity gain at all.

This upends the popular narrative that AI is a great equaliser, that anyone, even a newcomer with an AI assistant, can do the work of a veteran. In a perspective published in the same issue of Science, Lingfei Wu of the University of Pittsburgh and Bogdan Vasilescu of Carnegie Mellon University argue that AI does not lower the productivity bar. It raises it.

When generating code becomes cheap and easy, the bottleneck shifts from writing to judging. Developers must evaluate how well what AI produces actually works inside a complex system. Senior developers, drawing on years of experience with how software fails, can spot fragile suggestions. Junior developers often accept output that looks right in isolation but breaks under real-world conditions. Wu and Vasilescu call this a threshold model. Productivity gains kick in only after a minimum level of expertise.

This is what happens when the marginal cost of production collapses. Value migrates upward invthe skill chain, from routine execution to high-level design and judgment. Manufacturing went through this transformation decades ago. Automation reduced the need for routine assembly work and increased the premium on engineers and system architects. Something similar may now be unfolding in software and other knowledge industries.

A different investigation by MIT Technology Review, published in December, adds texture. After speaking to more than 30 developers, executives, and researchers, the publication found a more complicated picture than either the boosters or the sceptics portray. Developers agreed that AI excels at tedious, repetitive work. But for the complex problems where experienced engineers really earn their keep, the tools are not enough. Current AI models have limited working memory. They lose track of what they are doing on longer tasks and fail to account for how different parts of a large software system interact.

The implications are not confined to programmers. Wall Street is in the grip of what traders at Jefferies have christened the “SaaSpocalypse”, a massive sell-off in software stocks. The logic is simple. Most enterprise software companies charge per user, per seat. If AI agents can do the work of five employees, you no longer need five seats. You need one.

U.S. enterprise software stocks erased roughly $1 trillion in market value as investors reassessed how AI may compress demand for traditional software seats. The S&P 500 Software and Services index slid markedly, and companies such as Salesforce, Workday, and Adobe have declined along with it. The impact was visible in India as well. The Nifty IT index lost about 21%, and major firms including TCS, Infosys, HCL Technologies, and Wipro retreated.

Meanwhile, students are reading the signals. In the United States, computer science enrolment started to dip at many universities in 2025, the first broad decline since the dotcom bust. At the University of California system, enrolment dropped 6%. At Georgia Tech, new computer science enrolment fell between 2022 and 2025. The campus that bucked the trend was UC San Diego, which has a dedicated AI major.

The Delhi summit showcased India’s ambitions. The research published the same week suggests that ambition alone does not determine who benefits in an AI world. When the marginal cost of routine cognitive work collapses, the most productive, deeply skilled individuals become more valuable.

The AI race is not just about adoption. It is about who is positioned to obtain value from it.