US programmer job growth nearly halved since ChatGPT launched, Fed study finds

To strip out these industry-wide effects, the researchers built a counterfactual employment curve. It shows how many programmers there would be if their share within each industry had stayed constant and only the size of the industries had changed.

Even after this adjustment, programmer employment is still falling by about three percentage points a year. It looks like companies are deliberately cutting the share of programmers in their workforce. A control test using occupations barely touched by AI shows no comparable dip.

Half a million fewer jobs, but not necessarily job losses

Stretched over three years, the gap works out to roughly 500,000 jobs that probably would have existed without the rise of large language models. But the authors strongly caution against reading this number as a straight count of lost jobs.

Many would-be programmers likely found work in adjacent fields. AI may be reshuffling tasks across job categories, with programming skills now showing up more in other roles. The study also doesn't capture broader macroeconomic feedback effects. If AI lifts productivity broadly enough, total demand for labor could even grow over the long run.

The researchers also found no clear drop in wages. So far, the effect has shown up mainly in the number of jobs filled rather than in pay. Data from Indeed shows that job postings for software developers have been largely stable since 2024 and have ticked up slightly in recent months. Before that, they had fallen by more than half in 2022 and 2023.

The typical programmer isn't in Silicon Valley

One finding concerns industry structure: around 40 percent of all US programmers don't work at major tech companies or startups but at IT service providers that build software on contract. This contract sector is by far the biggest employer of programmers, and it's where the slowdown is most pronounced.

According to the study, the gap between actual and expected employment didn't open up until mid-2024 - roughly 1.5 years after ChatGPT launched. If AI is behind the shift in hiring plans, companies seem to have needed time to see the models improve before pulling back. The data doesn't make clear whether real productivity gains or just the expectation of them drove the decision.

AI may not be the only factor at play. A provision of the Tax Cuts and Jobs Act of 2017, which took effect in 2022, requires companies to amortize research expenses over several years instead of deducting them right away.

Since software development counts as research for tax purposes, this may have cooled hiring. But the existing research on its actual impact is mixed, and the study's results hold up even in sectors where the tax change should matter less.

The authors concede that pinning down a causal effect is tough with so many factors in play. They validate their approach using past disruptions: for bank tellers, the method correctly flags ATMs as an occupation-specific shock even as the banking industry as a whole grew. For seamstresses, it correctly identifies an industry-wide shock from textile manufacturing moving overseas.

Researchers don't agree on which jobs AI hits hardest

The study also points to a basic measurement problem in this field. There's still no standard way to figure out which occupations are most affected by generative AI. When it comes to identifying the group hit hardest, only about half of the common measurement approaches line up with one another.

Programmers are an exception, though: more than 98 percent fall into the most-affected group across every measurement method. That tracks with data from the Anthropic Economic Index, which shows that programming-related queries account for more than a third of all interactions with the chatbot Claude.

Whether the trend reverses over time as cheaper programming services open up new markets, whether offshore workers take a bigger hit, or whether AI ends up reshaping completely different occupations is still up in the air. The authors describe their work as "only a first step" toward answering these questions.

A recent study from Carnegie Mellon and Stanford found that AI agent development is almost entirely focused on programming tasks, while more economically significant areas barely register.