Why do we assume AI will take someone else's job first?

← Back to Blog

“People rating their own circumstances more favorably than other people's.”

Anthropic Economic Index (2026)

Ask a room of knowledge workers whether AI will eliminate their own job in the next two years, and most say no. Ask the same room whether AI will eliminate the job of the person next to them, and the number climbs. This is not hypothetical. Anthropic asked exactly that question and found the same pattern (Anthropic, 2026) (Emerging). In the same window, hiring of 22-to-25-year-olds into the most AI-exposed occupations slowed measurably against a modeled counterfactual (Massenkoff & McCrory, 2026) (Emerging).

Same company, same period, two numbers pointing in opposite directions. One is what employees say about their own risk. The other is what employers actually do when they staff entry-level roles. My hypothesis: organizations that use employee sentiment as their primary gauge of AI risk are reading a miscalibrated instrument, not a measured one. The gap between what people report and what hiring data shows is suggestive of that miscalibration, not proof of it, and the size of the gap is what deserves scrutiny.

The Research

Human-AI collaboration rests on the design of organizational structures, workflows, and decision rights that shape how people and AI systems work together, and that design discipline applies as much to labor-market measurement as it does to day-to-day work. The question is the same either way: who decided which instrument gets read, and does it actually track the thing they think it tracks?

Start with hiring, because it is behavior rather than forecast. Massenkoff and McCrory (2026), in Anthropic's own labor-market working paper, tracked hiring of 22-to-25-year-olds into the occupations most exposed to AI and found it had slowed relative to a counterfactual baseline (Emerging). That is hiring behavior, not sentiment. It records decisions employers made, not forecasts anyone offered. An independent estimate from Brynjolfsson, Chandar, and Chen (2025) puts the effect at a 6 to 16 percent fall in employment for that age group in exposed occupations (Emerging). Two analyses landing in the same range move this closer to established confidence than a single working paper would justify on its own, though the underlying evidence remains Emerging rather than Established.

Both measures share a boundary worth stating plainly: neither paper has completed peer review, and both remain working papers subject to revision. Massenkoff and McCrory (2026) build their counterfactual by modeling expected hiring absent AI exposure, then comparing observed postings against that model, a design that depends on how occupations get classified as AI-exposed and how far back the trend window extends. Brynjolfsson, Chandar, and Chen (2025) use a different exposure taxonomy and a different data source. Convergence between two independent designs is meaningful, but it is methodological triangulation, not replication in the strict sense the term earns in psychology, and the distinction matters for how much weight either number should carry on its own (Emerging).

Now the survey side. Anthropic's June 2026 Economic Index asked workers to forecast their own job-loss risk against their peers' risk and found systematic self-favoring, describing it as "people rating their own circumstances more favorably than other people's" (Anthropic, 2026) (Emerging). This is not new psychology. Festinger (1954) established that people calibrate self-assessment through comparison with others (Established), Weinstein (1980) documented the specific asymmetry in which individuals rate their own future risk below the population base rate for identical events (Established), and Miller and Ross (1975) trace the same self-protective pattern back to attribution research in organizational psychology (Established). Anthropic's survey is a fresh instance of an old finding.

What This Means in Practice

Picture a professional services firm reviewing its annual engagement survey. Junior analysts report low worry about AI eliminating their roles. Leadership reads that as reassurance. But the same firm's own recruiting funnel for those exact roles has quietly shrunk two quarters running. The sentiment data says calm. The requisition data says contraction. If the firm only looks at the survey, it misses the signal sitting in its own applicant tracking system.

The fix is not complicated, but it requires pulling data the organization already has into the same room. A workforce planning team should compare entry-level requisition volume in AI-exposed job families against sentiment survey results, the same comparison Anthropic makes at economy scale, run instead at company scale. Complexity theory offers a useful lens here: hiring, sentiment, and role redesign feed back on each other rather than sitting as independent readings (Anderson, 1999) (Speculative). A firm that redesigns roles in response to a hiring slowdown changes the very sentiment it will measure next quarter, which means a single snapshot from either instrument is always partial.

Three Things to Take Away

Pair sentiment with a behavioral metric

Do not trust an engagement survey on AI risk without a companion number pulled from actual staffing behavior. Entry-level hiring in exposed roles is one candidate, and Massenkoff and McCrory (2026) built exactly this kind of measure as a check against sentiment data alone.

Build feedback loops that surface peer data

Self-assessment gaps close when people see comparative evidence, not just their own forecast. Festinger's (1954) social comparison theory is the mechanism: exposure to how others rate the same risk recalibrates the individual's own estimate.

Treat any single data source as a hypothesis

Even the labor-market analysis you trust most should get checked against an independent dataset before it drives a decision. Brynjolfsson, Chandar, and Chen (2025) arrived at a comparable estimate using different methods, and Massenkoff and McCrory (2026) explicitly frame their own measure as provisional, meant to be updated by future observers.

My Two Cents

Here is what troubles me about how most organizations gauge AI risk, and it runs deeper than method. Feelings are data, but they are a specific kind of data, shaped by the same self-favoring bias Weinstein (1980) documented in risk perception more broadly, and psychology has known this since Festinger (1954) described comparison as the mechanism of self-assessment. A pulse survey on AI risk is not a weak instrument in the ordinary sense. It is a well-characterized one, with a predictable direction of error built into the question itself.

If a company's entire AI-risk measurement strategy is a pulse survey, it has instrumented its human-AI collaboration with exactly one sensor, and that sensor is known to run optimistic. I do not think this is a case for ignoring sentiment. I think it is a case for refusing to let sentiment stand alone.

There is a larger question sitting underneath any single firm's dashboard. The hiring slowdown Massenkoff and McCrory (2026) measure lands first on workers age 22 to 25, the group with the least tenure, the least negotiating leverage, and the fewest institutional protections in most organizations (Speculative). Nothing in the Anthropic survey breaks out self-favoring bias by age, so I cannot claim this group misjudges its own risk more than any other group does. What I can say is that if the same bias holds here too, it compounds badly: the workers most exposed to an actual hiring slowdown would also be the workers least likely to see it coming in their own sentiment data. The hiring data and the survey data disagree, and that disagreement matters most for the group whose outcomes are hardest to reverse once the pattern shows up in the numbers.

Read to Learn More

Weinstein, N. D. (1980). Unrealistic optimism about future life events. Journal of Personality and Social Psychology, 39(5), 806-820. The originating study on comparative optimism bias, still the standard citation for why self-ratings of risk diverge from population base rates.

TechInformed. (2026). Anthropic paper finds no AI job shock, but hiring slows. A concise industry read on the same Anthropic data, useful for readers who want the headline findings without the full working paper.

References