Anthropic maps real-world AI job risk: Programmers top list at 75%

For the past two years, AI labor market studies have mostly swung between doomsday scenarios and techno-optimist hand-waving. Everyone has an opinion on how many jobs will vanish, but hard proof has been remarkably thin.

Anthropic just dropped a new research paper that cuts through the noise. Instead of just theorizing about what large language models can do, they cross-referenced theoretical capability with actual telemetry from Claude to measure what they call observed exposure.

Theory meets the real world

Previous benchmarks often relied on a simple formula: if an LLM can theoretically cut the time required for a task in half, that task is considered exposed to AI.

Under that theoretical lens, roughly 94% of tasks in Computer and Math occupations are ripe for automation. Office and administrative roles sit right behind at 90%.

Anthropic’s observed exposure metric grounds those numbers in reality. They tracked how tasks actually appear inside Claude traffic, giving full weight to fully automated API workflows and half weight to augmentative, human-in-the-loop usage.

The resulting gap is massive. While 94% of computer and math tasks are theoretically feasible for an LLM, Claude currently covers only 33% of them in practice.

Why the lag? Theoretical capability hits walls like compliance hurdles, integration friction, and the need for human verification. An LLM could theoretically authorize prescription refills, but almost nobody lets an AI do that unsupervised today.

Still, some roles are already deeply in the crosshairs. Computer Programmers topped the list with 75% observed exposure, followed by Customer Service Representatives and Data Entry Keyers at 67%.

On the flip side, 30% of the American workforce sits at absolute zero exposure. If you are a cook, motorcycle mechanic, lifeguard, or bartender, your daily routine has not registered in Claude's logs at all.

The demographics of the vulnerable are also striking. Workers in the most exposed quartile earn 47% more on average, skew female, and are far more educated. Workers with graduate degrees make up 17.4% of the most exposed group, compared to just 4.5% of the unexposed cohort.

The subtle freeze on junior talent

If AI is taking over tasks so quickly, why haven't white-collar unemployment lines exploded?

Anthropic analyzed employment data since late 2022 and found no systematic spike in unemployment for highly exposed workers. The disruption is showing up differently.

First, future growth expectations are slowing. Data from the Bureau of Labor Statistics shows that for every 10 percentage point increase in observed exposure, projected 10-year job growth drops by 0.6 percentage points.

Second, and far more concerning, is a noticeable slowdown in the hiring of younger workers within highly exposed roles.

Companies aren't necessarily firing their experienced staff. They simply aren't posting as many entry-level openings.

Why it matters

We think Anthropic’s framework is one of the most honest assessments of the AI job market yet. It separates model parlor tricks from actual enterprise adoption, proving that capability alone does not equal immediate displacement.

However, the real danger here isn't a sudden wave of layoffs; it is a structural hiring freeze. For junior developers, entry-level customer reps, and recent college grads, the ladder is quietly being pulled up.

When experienced employees use AI to double their output, companies do not need to backfill entry-level slots. The dirty work that used to train the next generation is now being routed directly to APIs.

There is a blind spot, of course: this study reflects Claude’s ecosystem, leaving out proprietary enterprise deployments of OpenAI or open-source weights. But the direction is unmistakable. Don't wait for unemployment numbers to spike before updating your skills—by the time the macro data catches up, the entry-level door will already be closed.