The Roles That Use AI the Most Feel the Most Threatened — The Asymmetry Revealed by Anthropic’s Report

The same person says “AI has made me far more productive” and “my job is at risk” in the same breath. Cognitive dissonance, or rational judgment? Anthropic’s latest economic report leans toward the latter.

The Same Tool, Opposite Feelings

On April 23, 2026, Anthropic released a survey of AI’s economic impact based on its own user data. Gigazine summarized the report in Japanese on the same day, and its core finding compresses into a single line: “the more a role uses AI, the higher its perceived threat to employment.”

It sounds counterintuitive. The standard hypothesis would be this: people who use AI well are people who have tamed it as a tool, and therefore should feel less threatened. The ones who feel threatened should be those who cannot use it. The actual data points in the opposite direction. The higher the role’s exposure intensity, the stronger the perceived threat.

This pattern is hard to attribute to mere statistical noise. The mean productivity gain Anthropic measured is 5.1 on a seven-point scale. That means a majority of respondents are experiencing “major productivity gains.” Yet those same respondents are simultaneously reporting threats to their jobs. They are holding a large gain in one hand and a large anxiety in the other.

This article analyzes what this asymmetry means and what implications it has for enterprise decision-makers — those who issue outsourcing contracts or push internal AI adoption. The conclusion, stated up front, is that this data is neither a reason to push adoption hard nor a reason to hold it back. It does, however, demand a more precise answer to the question: “who, how, and at what speed?”

Five Patterns the Report Reveals

Pattern 1: A Positive Correlation Between Exposure Intensity and Perceived Threat

Anthropic constructed an “exposure” metric for each respondent that measures Claude usage intensity. The higher this metric for a given occupational group, the higher their perceived probability that AI will affect their own jobs in the future. In statistical terms, a positive correlation between exposure and perceived job threat.

This result admits two simultaneous interpretations. One: “those who have used AI heavily and seen its capabilities up close know how far AI can go.” What looks like a vague possibility from outside is, for the user, concrete reality. Half of the tasks the respondent personally handles with Claude each week were listed as core duties in their job description five years ago. That fact suggests two things at once: an expansion of their own capability, and a dilution of their own value.

Pattern 2: Differences by Career Stage

In the same report, Anthropic also analyzed career-stage variables. The result was clear: early-career professionals — i.e., junior-level workers — reported job-loss concerns significantly higher than senior workers did.

This is consistent with other data. Salesforce hired zero new SWEs during the year 2025. Computer-science enrollment at U.S. universities has been reported to have fallen by as much as 62% from its peak. The hypothesis that AI is reducing junior-engineer positions faster than it is replacing senior engineers is gaining ground in industry.

Anthropic’s new data suggests that “juniors’ concerns are not vague pessimism but perception grounded in data.” It is not simply “they’re young so they’re more afraid”; rather, they may be reading market signals first.

Pattern 3: Imbalance in Income and Benefits

The distribution of productivity gains was not uniform across roles. Management, software engineering, and other high-income roles captured the largest share of productivity gains. This partially supports the hypothesis that AI is closer to a tool that makes the already-skilled even more skilled than to a leveling tool that lifts everyone equally.

In economics this is called “skill-biased technical change”: a pattern in which new technology raises the value of high-skilled labor and lowers that of low-skilled labor. The effects of computerization since the 1990s are a canonical example. One reading of this data is that AI is repeating the same pattern, more forcefully.

Pattern 4: “Expanded Scope” and “Speed Increase” Are 90%

Nearly 90% of the gains respondents reported clustered in two categories.

First, expanded scope — the effect of enabling work outside one’s own specialty. Non-engineers writing code; an HR staffer handling SQL; a designer touching front-end code directly. Second, speed increase — one respondent reported “a finance task that used to take two hours now takes 15 minutes.” That is an 8x speedup.

On the surface, both patterns are positive. But on closer look, they carry a double edge. Expanded scope also means the boundaries of one’s own role are blurring. If a front-end engineer can write back-end code directly, then from the back-end engineer’s perspective, the number of people entering their territory is increasing. An 8x speedup also implies an 8x increase in throughput per person — that is, the same work needs roughly eight times fewer people.

Pattern 5: Maximum Threat at Both Extremes

The most interesting pattern is at the end. At both extremes of speed change — those whose speed dropped sharply and those whose speed rose sharply — perceived job threat was highest.

Why a large speed gain links to threat is intuitive: “if it gets this fast, fewer people will be needed.” But the threat felt by the slowed-down group tells a different story. These are people who adopted AI but whose workflows broke, whose review and correction costs increased, or for whom output unreliability extended their total time. Even so, the very fact that “we adopted AI” is read by management as efficiency pressure, raising the worry that actual efficiency fell while headcount could still be cut.

A pattern from the creative sector adds to this. According to the report, creative roles adopt AI more slowly than other sectors but perceive threat very strongly. This reflects the rational concern that AI can push down market prices even where it does not directly penetrate. If AI-generated illustrations and copy lower market rates, you are affected even if you do not use AI yourself.

Why the Same Person Holds Both Feelings at Once

The pattern of the same respondent simultaneously reporting productivity gain and job threat looks like a contradiction at first. But it can be rationally decomposed.

First, a difference in time horizons. Productivity gain is what you feel right now. Job threat is a possibility one to three years out. The two feelings are looking at different time axes, so there is no contradiction.

Second, the separation between personal utility and market value. That you become 8x faster with AI is personal utility. At the same time, that 8x means that your hourly market rate may become one-eighth, because AI strengthens you while commoditizing your unique value.

Third, the result of rational judgment. The cognitive dissonance hypothesis treats holding two contradictory feelings simultaneously as irrational. But given the two factors above, the pattern in this data is closer to evidence that respondents are reading their own situation very accurately. The more someone uses AI, the more accurately they measure the distance between AI’s capability curve and their own.

The Report’s Limits and Conflicts With Other Data

This Anthropic survey carries specific sample conditions. The respondents are Claude individual-account holders who self-selected into responding. This creates a possibility of self-selection bias. Those who have personally benefited most from AI may be more inclined to respond, and the 5.1/7 mean uplift may therefore be overestimated relative to the true population.

The more important contradiction is the conflict with other surveys from the same period.

According to the firm-level survey released by U.S. NBER in February 2026, approximately 80% of companies that adopted AI did not report meaningful productivity gains. Another survey released by MIT in 2025 concluded that 95% of enterprise AI pilots did not produce ROI. Gallup’s measurement of Gen Z sentiment toward AI in the same period showed excitement falling from 36% to 22% and anger rising from 22% to 31%.

How should we read this contradiction — that individual respondents report large gains while firm-level measurement detects no gain? Three possibilities.

First, self-reporting bias. Individuals tend to overestimate the time they saved and underestimate the added review and correction time.

Second, workflow effects not captured. Even when a step gets faster, if its output creates a bottleneck at the next step, total throughput does not rise. Even if a developer writes code 8x faster, if code review, QA, and deployment stay the same, shipping speed does not become 8x.

Third, self-selection bias. Individuals who responded are heavy AI users, while firm-level surveys include both heavy and light users.

All three hypotheses are likely partly true. Either way, the conclusion is the same: an individual respondent’s 5.1/7 does not guarantee firm-level ROI.

The Perspective of Clients and Internal AI Adoption Leads

Let us shift this data to the perspective of client firms — companies that take in outsourced IT labor or that recommend AI adoption to internal staff.

Verifying a Vendor’s Claim of “X% Faster With AI”

In quotes and scheduling discussions, more vendors are pitching that “we adopted AI and we handle the same work X% faster.” The implication of this report is that such claims cannot be taken at face value. Even when a 5.1/7 gain is reported at the respondent level, how that translates into project-level deliverable quality, schedule, and total cost is a separate question.

Worth examining:

Time measurement by stage. In which stage — coding, code review, QA, integration testing — did AI cut how much time?
Rework rate. How does the rework frequency of AI-produced artifacts compare with that of human-produced ones?
Staffing composition. Did headcount fall under the same schedule and quality bar, or did the same headcount produce more output?

If none of these three can be answered clearly, the claim “faster with AI” is closer to a qualitative impression.

Managing Internal Employees’ Sense of Threat

One of the most important implications of this report is that employees’ job concerns are not irrational pessimism but data-grounded perception. The fact that employees who use AI most actively are the most worried suggests that the harder a company pushes AI adoption, the faster employee concern can grow.

This cannot be eliminated entirely. But several approaches are worth examining.

Explicit role redefinition. The company should draw, in advance, where the core value of each role moves to after AI adoption, and share that picture with employees. The longer a company delays this, the more employee anxiety fills in with speculation.
Concrete retraining tracks. Not an abstract promise of “retraining” but concrete paths describing which skill set, in what timeframe, the transition runs through.
Separate design for early-career. In light of data showing junior concerns are largest, hiring policy and junior-role design require separate review. A strategy of simply “cut juniors, replace with seniors” risks breaking the talent pipeline over the medium to long term.

Caution in the “AI-Augmented Role” Layoff ROI Calculation

The most dangerous decision is the linear inference that “role X became faster with AI, so we can cut headcount.” This data suggests two things.

First, individual-level gains are likely not to translate into firm-level ROI (see NBER and MIT data). Second, the employees who feel most threatened are likely the ones already using AI best. Cutting them risks reducing the company’s AI capability along with them.

If layoff decisions are necessary, the calculation should be not “cut by however much faster we got with AI” but, first, “where do we redeploy the time we saved?” Until the latter is answered, the former is worth holding.

Scenarios: Short, Medium, Long

Three scenarios can be drawn from this data.

Short term (1 year): The reinforcement pattern in which those who use AI well use it even better dominates. Productivity gains concentrate in high-income roles including management and SWE, while juniors and creative roles feel more threatened. Firm-level ROI remains unclear.

Medium term (2–3 years): A bifurcation. On one side, companies that have converted individual gains into firm-level ROI through workflow redesign and role redefinition. On the other, companies that could not, and have only added cost. The gap between them widens. Role polarization deepens.

Long term (5+ years): Two possibilities. One: firm-level ROI is established and AI integration becomes fully routine. Two: the bubble deflates and AI adoption retreats in some areas. In either case, workforce composition and role definitions undergo one more reshaping.

All three scenarios share one thing in common: in the short term, the sense of threat is justified by data.

Conclusion: What the Asymmetry Asks of Us

Back to the lede question. Is the same person reporting “large gains from AI” and “threat to my job” cognitive dissonance, or rational judgment? The data in this report leans toward the latter. The fact that a tool that strengthens you can simultaneously commoditize you is most accurately read by those who use it most.

This is not a conclusion that AI adoption should stop. Nor that it should be accelerated. It is a signal that the moment has come for the whole company to hold a more precise answer to the question of “who, at what speed, and with what safeguards.”

For IT managers, HR, and executives at client firms, the question this data poses is simple. If the employees who use AI best at our company are also carrying the greatest anxiety, in what direction does an adoption strategy that fails to read that anxiety pull our company? Deciding adoption speed before this question has an answer is to ignore the asymmetry the data suggests.

The asymmetry cannot be ignored. It will bill us for its cost in next quarter’s workforce plan.

Sources

Anthropic AI Economic Impact Report (2026-04-23)
Gigazine coverage: Correlation between AI usage intensity and perceived job threat (2026-04-23)
NBER: Enterprise AI Productivity Survey (2026-02)
MIT: Enterprise AI Pilot ROI Study (2025)
Gallup: Gen Z AI Sentiment Tracker (2026-03)