Anthropic and Stanford studies clash over how big AI’s hit on programming jobs will be

Programming jobs are becoming the test case for the whole AI labor debate. Not because coders are disappearing overnight — they are not — but because software work sits right where large language models are strongest. That is why two closely watched studies are getting so much attention now. Anthropic is arguing that programming is the most AI-exposed occupation in its real-world usage data, while Stanford economists are seeing the first clear labor-market damage show up in younger workers and hiring pipelines. (anthropic.com) ### What did Anthropic actually say? Anthropic’s March 5, 2026 paper built a measure it calls “observed exposure.” Basically, it tries to combine what frontier models can theoretically do with what people are actually using them to do at work. In that framework, programmers land at the top of the list, with 74.5% observed exposure. But Anthropic’s own catch matters — actual AI coverage(anthropic.com)ays large-scale unemployment has not yet shown up in the data. (anthropic.com) ### What did Stanford actually find? The Stanford paper is looking at labor-market outcomes, not model capability. Using ADP administrative payroll data, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen found that workers ages 22 to 25 in AI-exposed occupations saw a 16% relative employment decline after generative AI spread widely, while more experienced workers in those same occupati(anthropic.com)y — which is another way of saying the first hit seems to be fewer seats, not lower wages. (digitaleconomy.stanford.edu) ### Why do these studies sound like they disagree? Because they are measuring different things. Anthropic is asking where AI can substitute for work, especially when people use it in automating ways instead of assistive ones. Stanford is asking whether that substitution has already changed headcount. Those are related, but they are not the same question. A job can be highly exposed before the labor market fully reprices it. (anthropic.com) ### Why are young developers the first place to look? Entry-level software work contains lots of the stuff AI is already good at — boilerplate code, debugging help, test generation, documentation, and translating specs into first drafts. That does not eliminate senior engineers. If anything, it can make experienced people more productive. But it can reduce the number of juniors a team (anthropic.com)ly: employment for software developers ages 22 to 25 has fallen nearly 20% from 2024. (hai.stanford.edu) ### So are programming jobs shrinking overall? Not in the simple, headline way people keep reaching for. Federal projections still show software developers, QA analysts, and testers growing 15% from 2024 to 2034, with about 129,200 openings a year on average. More broadly, computer and IT occupations are still projected to grow faster than the economy-wide average. So the cur(hai.stanford.edu)n the floor above it.” (bls.gov) ### What should people take from this? The biggest mistake is treating exposure as displacement. Anthropic’s 74.5% figure does not mean three-quarters of programmers lose their jobs. But the other mistake is pretending nothing is happening because total software demand still looks healthy. The early damage seems concentrated in junior hiring, first jobs, and transition points into the field. (anthropic.com) ### What is the real bottom line? Both studies can be right at once. Programming may be the occupation where AI touches the most tasks, and the first labor effect may still be narrow — fewer young developers getting in, not mass layoffs of everyone already there. That is a smaller claim than the doom version, but it is also more immediate. And for new grads, it may be the part that hurts first. (anthropic.com)