Anthropic Releases Early Warning System for AI Job Exposure

Here's a question that's been impossible to answer properly until now: How much is AI actually changing the job market, versus how much could it theoretically change things?

Turns out, there's a massive difference between those two numbers. And Anthropic just built a system to track it.

On March 5, 2026, Anthropic economists Maxim Massenkoff and Peter McCrory published research introducing what they call "observed exposure" — a new metric that doesn't just measure what AI could do to jobs, but what it's actually doing right now. The findings are both reassuring and quietly alarming, depending on where you look.

What Anthropic Actually Built (And Why Previous Measures Fell Short)

Every major AI jobs study until now has had a fundamental flaw: they only measured theoretical capability.

The most influential prior framework, from OpenAI's Eloundou et al. (2023), scored tasks on a simple scale — could an AI double the speed of this task? Yes or no. That became the standard metric across dozens of studies. But as analysis from Development Corporate points out, theoretical capability doesn't equal economic displacement.

Here's a perfect example of the problem: the Eloundou framework classified "Authorize drug refills and provide prescription information to pharmacies" as fully AI-exposed. According to Development Corporate's analysis, Claude has reportedly never performed this task at scale — despite the technical feasibility. Legal constraints, liability requirements, and verification workflows create a deployment gap that purely theoretical models simply can't see.

Anthropic's new approach combines three data sources:

• O*NET — the U.S. Department of Labor database cataloguing tasks across 800+ occupations

• Anthropic's own Economic Index — anonymised, real-world Claude usage data from professional interactions

• Eloundou et al.'s theoretical feasibility scores — used as a ceiling measure

The methodology according to Anthropic's paper weights automated use at full weight over augmentative use (half weight), and only counts a task as "covered" if it shows sufficient work-related usage in actual Claude traffic. Not just whether an AI could do it — whether people are actually using it for that.

The Top 10 Most Exposed Professions

So which jobs are feeling the most real-world AI impact right now? According to CBS News's reporting on the study, here's the list:

• Computer programmers: 75% of tasks covered

• Customer service representatives: 70%

• Data entry keyers: 67%

• Medical record specialists: 67%

• Market research analysts and marketing specialists: 65%

• Sales representatives: 63%

• Financial and investment analysts: 57%

• Software quality assurance analysts: 52%

• Information security analysts: 49%

• Computer user support specialists: 47%

At the other end of the spectrum? Around 30% of the workforce has essentially zero exposure. Groundskeepers, cooks, motorcycle mechanics, lifeguards, bartenders — jobs requiring physical presence that CBS News reports ranked among the lowest exposure levels.

The Capability Gap: What AI Could Do vs. What It Actually Does

This is where it gets really interesting — and where the nuance lives.

According to reporting from Tech in Asia, in computer and maths occupations, AI is theoretically capable of performing 94% of tasks. But actual observed Claude usage covers just 33%. That's a staggering gap.

Office and administrative roles show a similar pattern — according to Infobae's coverage, theoretical exposure sits at 90%, but real-world usage is a fraction of that. Legal occupations are theoretically highly exposed but see minimal actual AI deployment, largely due to liability constraints.

Now, should that gap make you breathe easy? Not necessarily. As the Anthropic researchers note in their paper, "AI is far from reaching its theoretical capability: actual coverage remains a fraction of what's feasible." That theoretical ceiling isn't getting lower — and as implementation deepens, that gap will narrow.

As Infobae pointedly noted: "What today is a gap is, in reality, a deadline."

No Mass Layoffs — But Something Quietly Concerning Is Happening

Here's the headline that will make some people relax prematurely: according to Anthropic's research, there has been "no systematic increase in unemployment for highly exposed workers since late 2022."

But don't stop reading there.

The study found according to Development Corporate's analysis "suggestive evidence" of a 14% drop in job-finding rates for young workers aged 22–25 entering high-exposure occupations. Companies aren't laying off experienced staff. They're quietly freezing entry-level hiring.

That distinction matters enormously. Mass layoffs make headlines. A slow, invisible reduction in entry-level opportunities doesn't — but it shapes an entire generation's career trajectory.

The Register notes that even the researchers themselves were cautious about this finding, describing the 14% figure as a "suggestive" average estimated effect rather than a definitive one. But the pattern is there — and it's worth watching closely.

The Demographics Might Surprise You

If you're picturing factory workers or low-wage roles when you think of AI job exposure, think again.

According to Anthropic's research, workers in the most AI-exposed roles are:

• 16 percentage points more likely to be female (54.4% vs. 38.8% in non-exposed roles)

• Earning 47% higher average wages than workers in non-exposed jobs

• Nearly four times more likely to hold a graduate degree (17.4% vs. 4.5%)

• More likely to be older than workers in non-exposed occupations

This flips the traditional automation narrative on its head. We're not talking about assembly lines — we're talking about the professional class. As CBS News notes, this aligns with previous research finding that women-dominated occupations like administrative assistants and clerks are particularly vulnerable.

The Awkward Contradiction Anthropic Probably Doesn't Love

Here's something worth flagging: there's a notable tension between this research and Anthropic's own leadership.

As The Register highlights, Anthropic CEO Dario Amodei reportedly predicted in January 2026 that "AI could displace half of all entry-level white collar jobs in the next 1–5 years." His own economists' measured findings? "No systematic increase in unemployment" and effects "indistinguishable from zero."

That's a pretty stark internal contradiction. It doesn't invalidate the research — if anything, the fact that Anthropic published findings that contradict its CEO's public statements lends the study some credibility. But it's a reminder to take any single company's framing with a healthy grain of salt.

What We Don't Know Yet

This research is genuinely valuable, but it has significant blind spots we should be honest about:

• It only covers Claude. Anthropic's "observed exposure" metric relies entirely on its own model's usage data. We don't know if usage patterns for ChatGPT, Gemini, or specialised enterprise AI tools differ significantly — and they probably do.

• The youth hiring slowdown needs more data. A 14% drop in job-finding rates for 22–25 year olds is concerning, but the long-term compounding effects — on management pipelines, career development, economic stability — are completely unmapped.

• There's no timeline for when the gap closes. The study documents the chasm between theoretical and actual AI deployment, citing legal constraints, model limitations, and human review requirements. But it doesn't forecast when those hurdles will be cleared.

• The broader economic picture is messy. Infobae reports that the U.S. economy lost 92,000 jobs in February 2026, with the information services sector losing 11,000 jobs. Isolating AI's specific role from broader economic forces remains extremely difficult.

What This Means for You — Practical Takeaways

Whether you're a programmer, a marketer, a manager, or a new graduate, here's how to think about this:

If you're in a high-exposure role: Don't panic. The data shows you're not about to be replaced tomorrow. But start thinking about which parts of your job AI handles well versus which parts require judgment, creativity, and human relationships. Lean hard into the latter.

If you're early career or a new graduate: This is the finding that should sharpen your focus. Entry-level roles in AI-exposed fields appear to be quietly contracting. Build demonstrable AI skills, seek roles that combine human judgment with AI tools, and consider positioning yourself as someone who can work with AI rather than in spite of it.

If you manage teams or make hiring decisions: The capability gap won't last forever. Now is the time to develop transition plans for roles that sit high on the exposure list — not in a year when the gap has narrowed.

If you're evaluating AI companies' claims: This research is a powerful reminder that what AI can do and what it is doing are wildly different things. Be sceptical of vendors claiming wholesale automation of complex workflows.

The Bigger Picture: An Early Warning System, Not a Verdict

Here's what I genuinely appreciate about this research: Anthropic's researchers explicitly frame this as establishing a baseline to revisit periodically, "before meaningful effects have emerged," so that future analyses can more reliably identify economic disruption than post-hoc studies.

That's a refreshingly humble approach. As the researchers note, "the track record of past approaches gives reason for humility." Previous attempts to predict job disruption — from offshoring to industrial robots — have often gotten it wrong. This framework is designed to catch the signal before it's overwhelmed by noise.

The occupations with higher observed exposure are projected by BLS data to experience slower employment growth through 2034. That's not a catastrophe — but it is a clear, data-backed signal about where the labour market is heading.

Quick Takeaways

• The gap is real and enormous: AI can theoretically do 94% of computer/maths tasks but is actually being used for just 33%. The distance between capability and deployment is your window to adapt — but it's a window, not a wall.

• No mass layoffs yet, but watch the entry level: Experienced workers aren't losing jobs, but there's suggestive evidence of a 14% drop in job-finding rates for young workers (22–25) in exposed fields. The disruption is happening through hiring freezes, not pink slips.

• The most exposed workers aren't who you'd expect: They're more educated, higher-paid, and more likely to be female. This is white-collar, professional-class exposure — and it's time to start treating it that way.