Why High-Paid, Educated Workers Face the Most AI Risk in America

The automation story everyone told - and why it was wrong for AI

The dominant narrative about automation and jobs over the past 50 years went like this: automation hits the bottom of the labour market first. It targets repetitive, physical, low-skill work. Factory workers, assembly line operators, and manufacturing employees face the wave while office workers, professionals, and knowledge workers are insulated by their education and earnings.

That story was accurate for decades. The automation of the 1970s through 2010s - industrial robots, computerised manufacturing, logistics automation - did follow that pattern. It hollowed out the middle and lower-middle of the labour market while leaving the top relatively untouched.

AI is doing the opposite. And Anthropic's 2026 research, which measures actual Claude usage across approximately 800 US occupations, makes this visible with hard data for the first time.

The numbers: what Anthropic's data shows about income and exposure

Massenkoff and McCrory (2026) compared workers in the top quartile of AI exposure - those whose jobs are most actively being performed by AI in real workplaces today - against workers with zero observed AI exposure. The demographic gap between those two groups is striking.

Source: Massenkoff, M. and McCrory, E. (2026), Labor Market Impacts, Anthropic Research, March 5, 2026. Data measured from US worker demographics August-October 2022 (pre-ChatGPT baseline).

These are not small differences. A 47% earnings premium is enormous. A 4x difference in graduate degree prevalence means the high-exposure group looks nothing like the low-exposure group educationally. The data is measuring two very different segments of the US workforce - and the more educated, better-paid one is the one AI is targeting.

Why this happens: what AI actually does well

Understanding why AI targets high-earners requires understanding what AI is actually good at. AI - and specifically large language models like Claude - excels at three things:

• Language processing: Reading, writing, summarising, drafting, translating, editing. Any task where the primary input and output is text.
• Pattern recognition in structured data: Finding regularities, classifying information, applying rules consistently across large volumes of inputs.
• Reasoning from known information: Taking a set of facts or constraints and working through their implications - analysis, research synthesis, code generation from specifications.

Now ask yourself: who does those things at work? Not the cook, the mechanic, or the lifeguard. The lawyer drafting contracts. The accountant producing financial analyses. The software engineer writing code. The consultant synthesising research into a report. The data analyst producing models from spreadsheets. The HR manager writing job descriptions and handling correspondence.

High-paying knowledge work is mostly language, structured reasoning, and pattern matching. That is precisely the space AI occupies. The jobs that command premium salaries are premium-salary jobs partly because those skills were hard to find and hard to scale. AI removes both constraints simultaneously.

The gender finding: why women face more AI exposure than men

The gender dimension of AI risk is the one that most directly contradicts the old automation narrative. Industrial automation was predominantly a threat to male workers in male-dominated manufacturing sectors. AI exposure is 16 percentage points higher for women than for men in zero-exposure occupations.

This reflects occupational segregation in the knowledge economy. The high-exposure roles - administrative, clerical, HR, legal support, medical coding, data entry - are roles where women are heavily represented. The zero-exposure roles - trades, construction, mechanics, food service - skew male. AI is not designing this outcome deliberately. It is simply better at office work than at manual work, and office work has historically employed more women.

The implication is serious. If AI deployment continues on its current trajectory, the workers most affected in the near term will disproportionately be women in well-paying knowledge and administrative roles. That is a different policy challenge than the one governments designed their retraining programmes around.

The education paradox: why your degree may not protect you

Graduate degree holders are four times more prevalent in high-exposure occupations than in zero-exposure occupations. This is perhaps the most counterintuitive finding in the Anthropic data.

The standard human capital theory of technological change predicts that education protects workers from automation. Higher-skilled workers adapt, learn new tools, move into roles that automation creates. Lower-skilled workers bear the cost. This held roughly true for industrial automation - factory workers displaced, degree-holders largely unaffected.

AI breaks this pattern for a specific reason: graduate degrees cluster in knowledge work, and knowledge work is exactly where AI is most capable. A graduate degree in law does not protect you because AI cannot read. It makes you more exposed because AI can read very well - and reading, reasoning, and writing is most of what legal work requires. A master's degree in finance does not insulate you because financial analysis requires rare skills. It puts you squarely in the category of "tasks Claude is already doing in many workplaces."

The Asian American finding: concentration in tech and knowledge roles

Asian American workers are nearly twice as likely to be in high-exposure occupations as workers in zero-exposure roles. This reflects concentration in two high-exposure sectors: technology and professional services.

Asian American workers are overrepresented in software engineering, data science, financial analysis, and medical professions relative to their share of the overall labour force. All of these are high-exposure categories in Anthropic's data. This is not a finding about differential vulnerability within occupations - it is a finding about occupational concentration. The same logic applies as it does to the education finding: clustering in knowledge and technology roles means clustering in AI-exposed roles.

High pay is not the same as high protection

There is a common assumption that high earnings signal that a role is complex enough to resist automation. High pay, the logic goes, reflects the market's assessment that the work requires rare and difficult-to-replace capabilities. If the market pays someone well, automation must not have found a way to do their job yet.

Anthropic's data challenges this directly. The 47% earnings premium in high-exposure occupations tells you something important: AI is targeting jobs the market currently values highly. The reason those jobs pay well - they require language facility, analytical reasoning, and structured problem-solving - is precisely the reason AI can now do significant portions of them.

High pay buys you time in two ways. First, employers have more incentive to deploy AI carefully in high-value roles - errors are more costly, oversight requirements are higher, the savings from replacement need to justify the risk. Second, high-earning workers can afford to adapt, retrain, and move into adjacent roles. But high pay is not a structural barrier to AI adoption. It is a friction that slows deployment, not one that prevents it.

What the data says is actually protected

The zero-exposure group - the 30% of US workers with no meaningful observed AI usage in their occupations - is defined not by education or income but by physicality and unpredictability. Cooks. Mechanics. Lifeguards. Bartenders. What these roles share is not low skill or low pay in absolute terms. A skilled electrician earns more than many data-entry clerks. What they share is work that requires physical presence in environments that change constantly, where the inputs cannot be fully specified in advance and the outputs require hands and body, not just language and reasoning.

The protected jobs in Anthropic's data are protected for a structural reason: AI does not have a body. Until general-purpose physical robotics reaches the reliability and dexterity of human workers across variable real-world environments - which remains an unsolved problem in 2026 - physical roles maintain their structural protection. That protection is not guaranteed forever. But it is real now, and it is measurable in the data.

What high-exposure, high-earning workers should actually do

None of this means high-paid knowledge workers should panic. The deployment gap data shows that ICT professionals have a 61-point gap between theoretical and observed exposure - years of runway before the wave reaches its theoretical maximum even in the most-exposed high-earning category. But the direction is set, and the Anthropic data makes clear that high pay and high education are not the shields they were in previous automation cycles.

The practical response has three components. First, understand which parts of your role require things AI cannot yet do: novel judgment, accountability in ambiguous situations, physical client relationships, creative synthesis across domains with no established pattern. Lean into those. Second, become a user of AI tools in your own work rather than a competitor to them. The workers most likely to maintain their value in high-exposure fields are the ones who are directing AI, evaluating AI output, and handling the cases AI cannot. Third, pay attention to the hiring data for your field. The young worker hiring slowdown detected in the Anthropic research is the leading indicator - it shows up in who gets hired before it shows up in who gets laid off. If junior hiring in your sector is contracting, that is an early signal of what the automation wave is doing to the entry point of your career ladder.

Frequently asked questions