Anthropic's New AI Exposure Score: 22-25 Year-Olds Hit First

What if 75% of computer programmers' work is already being done with AI assistance, but their unemployment rate hasn't moved at all? That's not a hypothetical. That's the headline finding from a new Anthropic research paper that quietly dropped on March 5th, 2026 — and it might be the most important labor data of the year.

Here's why this paper matters more than the dozens that came before it. For three years, every "AI will replace X% of jobs" headline traced back to one 2023 paper by Eloundou and colleagues. That paper measured theoretical exposure — what AI could do in a perfect world. Anthropic's economists Maxim Massenkoff and Peter McCrory just released something fundamentally different: a measure of what AI is actually doing right now, inside real workflows, weighted by whether it's replacing the human or just helping them along.

They call it observed exposure. And the gap between theory and reality is the most reassuring thing you'll read all month — until you get to the part about young workers.

How "Observed Exposure" Actually Works

The trick is the data source. Anthropic has something no academic team had access to: anonymized records of how millions of Claude conversations map onto real occupational tasks. By matching those conversations to the O*NET task database covering roughly 800 U.S. occupations, the researchers built a per-job score that reflects what people are paying AI to do — not what economists guess it might one day do.

There's one more crucial wrinkle. When someone uses Claude to replace a task entirely (automation), it counts as full weight. When they use it to assist a task (augmentation, like brainstorming or editing), it counts as half weight. [Fact] That weighting matters because augmentation often makes workers more productive without removing them from the equation, while automation directly substitutes for hours of human labor.

The result is a scoring system that captures real-world adoption, software constraints, legal limits, and the simple question of whether anyone has bothered to point AI at a given task yet.

The 75% That Should Worry You — And the 33% That Shouldn't

The single highest-scoring occupation in the entire dataset is Computer Programmers at 75% observed exposure. [Fact] If you write code for a living, three out of every four hours of your job description, on average, are showing up in Claude conversations somewhere. That's the headline number every tech publication ran with.

But the more interesting number is buried one level deeper. The full Computer & Mathematical occupational category — which includes programmers, software developers, data scientists, and statisticians — has a theoretical exposure score of about 90%. Its observed score? Just 33%. [Fact]

That gap — the difference between what AI could touch and what it's actually touching — is the entire ballgame. It tells you that adoption, integration into existing software stacks, regulatory friction, and old-fashioned human inertia are doing more to slow the transition than any technical limit. The technology raced ahead. The workplace didn't.

For occupations like Customer Service Representatives and Data Entry Keyers, the gap is smaller — these are jobs where the software pipelines are simpler, the legal constraints lighter, and the integration cost lower. So adoption races ahead faster. If you're in one of those roles, the observed-exposure number isn't a forecast. It's a description of last quarter.

The 30% Who Are Effectively Untouched

Here's the part that gets buried in the doom narratives. About 30% of American workers have essentially zero observed exposure. [Fact] These are jobs where AI has not been measurably deployed at all — not because regulators stopped it, not because workers organized against it, but because nobody has figured out a useful way to apply it.

The list includes Cooks, motorcycle mechanics, lifeguards, Bartenders, dishwashers, and locker room attendants. The pattern is obvious once you see it: physical work, in-person presence, real-time judgment about messy environments. These jobs were supposedly the first to be automated, in every futurist prediction from the 1950s through the 2010s. They are the last now.

If you're in a hands-on trade, a service role, or anything that requires being in a specific physical place at a specific time, the data is telling you something important. The AI revolution is happening in front of screens, not in kitchens or repair bays.

The Young-Worker Signal That Demands Attention

Now the harder part. The paper finds that for every 10 percentage points of additional observed exposure in an occupation, BLS-projected employment growth drops by about 0.6 percentage points [Fact] — a small but statistically meaningful tilt. Across the whole labor market that's not huge. But it's not nothing, and it's concentrated in a way that should change how we read this data.

Where the signal sharpens is in new hires for workers aged 22-25. In high-exposure occupations, new hiring of this cohort has dropped by roughly 0.5 percentage points [Fact] in a way the researchers can isolate as statistically significant. For prime-age and older workers in the same fields, the effect disappears. Existing employees keep their jobs. The unemployment rate barely budges. The unemployment rate of high-exposure workers post-ChatGPT shows no significant increase compared to low-exposure workers. [Fact]

But the door for fresh entrants is closing — by a crack, not a slam, and not everywhere. This is exactly the pattern you'd expect if firms are quietly using AI to handle the entry-level tasks that used to be a junior employee's first six months on the job. The senior engineers don't lose their roles. But the company hires four new grads instead of five.

This deserves its own paragraph because it's the part of the paper that should change behavior. If you're 22-25, looking at a high-exposure field, you are competing against the same job posting your older colleagues filled five years ago — but with one fewer slot. The pipeline narrowed.

What This Means for Your Career

Here's the framing that I think the data actually supports — and it's more hopeful than the headlines will suggest. The AI exposure picture is not a prediction of mass unemployment. The data, in 2026, doesn't show that. What it shows is a slow rotation in the labor market, with three patterns:

1. High-exposure occupations have a real but modest hiring slowdown for the youngest cohort. The total wage gap between high- and low-exposure workers in 2022 CPS data was roughly 47% [Fact] — meaning these are still well-paid jobs. They're not disappearing. They're just hiring fewer fresh entrants.

1. The 30% with zero observed exposure are essentially untouched. Physical, in-person, judgment-heavy work is still firmly human. If you're worried about AI displacing you and you're in one of these roles, the worry is not supported by the data so far.

1. The augmentation half of the score is a feature, not a bug. When Claude is being used to assist rather than replace, it's making humans more productive — and that productivity is what funds the wages of the very workers using it.

The research disclosure deserves a quote: ChatGPT launched in November 2022. We are now three and a half years into the deployment. The unemployment rate of high-exposure workers has not significantly increased relative to low-exposure workers. [Fact] That is a remarkable fact and the headlines have not absorbed it yet.

What the headlines also haven't absorbed: this is not the world ending. This is the world adjusting. Workers in their 30s, 40s, and 50s in high-exposure fields are doing their jobs with new tools, getting paid roughly the same, and not getting laid off. The pressure point is precisely at the door — for new entrants — and that's a problem that's solvable with better internships, apprenticeships, and entry-level redesigns rather than panic.

Practical Advice if You're Reading This

For workers already in high-exposure fields: the paper is, frankly, mild reassurance. Your job is being augmented, not replaced. Learn the tools that augment it. Become the person who uses them best.

For young workers entering the labor market: take the 0.5pp hiring drop seriously but not catastrophically. The drop is real. It's also small. The way through is portfolio work, public projects, and a deliberate effort to demonstrate skills that aren't easily replicated by an AI prompt.

For workers in low-exposure fields: ignore the AI doom takes. The data does not say what those takes claim. Your work is not being automated. It might never be.

For everyone: keep watching this kind of research, not the press releases. Anthropic has effectively raised the standard for measurement in this field. Every "AI replaces X% of jobs" headline from now on should be checked against whether it's measuring observed exposure or theoretical exposure. They are different worlds.

---

This article was written with AI assistance and reviewed for accuracy. The underlying research, data, and conclusions are drawn from Anthropic's published economic research. Quotes are paraphrased; specific numbers are verified against the source paper.

Sources

• Massenkoff, M. & McCrory, P. (2026). "Labor market impacts of AI: A new measure and early evidence." Anthropic Economic Research, March 5, 2026. https://www.anthropic.com/research/labor-market-impacts
• Anthropic Economic Index, August and November 2025 releases
• O*NET task database (occupational task taxonomy)
• BLS Employment Projections, 2024-2034
• CPS (Current Population Survey) wage data, October 2022

Update History

• 2026-05-04 — Initial publication