Anthropic says AI displaces 74.5% of programmers

Programming work is where generative AI seems to be biting first. That is the real headline in Anthropic’s latest labor-market research — not that “all jobs are gone,” but that one company now has a concrete, usage-based way to estimate where AI is already doing meaningful chunks of work. And in that ranking, programmers sit at the top. The study landed on March 5, 2026, and it matters because it tries to measure actual use, not just hypothetical future capability. (anthropic.com) ### What did Anthropic actually measure? Anthropic built what it calls “observed exposure.” Basically, it mashed together three things: what large language models can theoretically do, how tasks map onto occupations in the U.S. O*NET database, and how people are actually using Claude in work-related conversations. The point was to move past the usual “AI could automate X” claims and anchor the estimate in real traffic. (anthropic.com) ### Why are programmers at the top? Because coding is already one of Claude’s heaviest real-world use cases. Anthropic’s earlier Economic Index work showed software development and writing dominating Claude usage, and the newer labor paper says computer programmers have about 75% observed exposure — reported more precisely elsewhere as 74.5%. That does not mean 74.5% of programming jobs va(anthropic.com)pation are now being touched by AI in a way that looks more automated than collaborative. (assets.anthropic.com) ### What does “observed exposure” mean in plain English? Think of it as task coverage, not pink slips. If a programmer uses Claude to write boilerplate, debug functions, explain stack traces, generate tests, or refactor code, those tasks count toward exposure. Anthropic weights automated uses more heavily than augmentative ones, so the metric is trying t(assets.anthropic.com)study got attention. (anthropic.com) ### Which other jobs showed up? Customer service representatives ranked next at 70.1%. Then came data entry keyers at 67.1% and medical records specialists at 66.7%. The pattern is pretty intuitive once you see it — jobs built around structured digital information, repetitive text handling, and standardized workflows are easiest for current models to absorb. (aigazine.com)ges-47-higher-in-toprisk-jobs--s)) ### Who looks most exposed? Anthropic says the most exposed professions skew older, more female, more educated, and higher-paid. That cuts against the older story that AI would mainly hit low-wage routine labor first. In this dataset, the early pressure is landing hardest on white-collar knowledge work — especially jobs that live inside documents, tickets, forms, spreadsheets, and code editors. (anthropic.com) ### Are jobs disappearing already? Not cleanly, at least not in the aggregate data Anthropic looked at. The paper says it does not find a systematic rise in unemployment for highly exposed workers since late 2022. But it does see hints that hiring for younger workers may be slowing in exposed occupations. So the immediate effect may be fewer entry points and leaner teams, not mass layoffs all at once. (anthropic.com) ### What’s the catch with this study? It is still Anthropic measuring usage of Anthropic’s own model. That makes the dataset unusually rich, but also narrow. It captures where Claude is used, not every AI tool in the economy, and it maps conversations onto occupations rather than directly observing payroll decisions. So this is best read as early evidence of task-level displacement pressure — strong evidence, but not the final word. (anthropic.com) ### So what should people take from it? The cleanest takeaway is that AI is no longer just a productivity sidekick for programmers. In some parts of software work, it is becoming the default first pass. The labor-market shock may arrive less like a sudden collapse and more like a hiring freeze that spreads outward — first in code, then in every job that looks enough like code for a model to handle. (anthropic.com)