For roughly three years, the labor market debate about AI lived entirely in the future tense. Economists published projections. Think tanks released reports with titles involving the word “transformation.” Both the catastrophists (40% of jobs gone by 2030) and the incrementalists (AI is just another general-purpose technology that creates as many jobs as it destroys) were operating on theoretical models dressed up as empirical claims. We’re now past the point where that dodge is available. The data for 2025 and early 2026 exists, and it’s time to read it without preferring one conclusion over another.

The headline finding is this: AI is causing real, measurable employment contraction in specific occupational categories, the contraction is significantly smaller than the catastrophist projections, it is significantly larger and faster than the incrementalists projected, and — crucially — the distributional pattern is exactly what should concern policymakers most.

Start with what’s contracting. The BLS Occupational Employment and Wage Statistics data for mid-2026 shows meaningful year-over-year employment declines (greater than 5%) in the following categories: legal support workers excluding paralegals, word processors and typists (a category that has been declining since the 1990s but accelerated sharply), credit authorizers and checkers, proofreaders and copy markers, and data entry keyers. These are not surprising. They are the roles that AI assistants handle competently on day one with minimal prompt engineering. The work is text-based, rule-bound, and legible to a language model.

What’s more interesting — and less reported — is the second-tier list. Roles experiencing smaller but statistically significant declines (2-4% year-over-year, which is notable against a background labor market that was adding roughly 1.5% annually): insurance underwriters, tax preparers below the CPA threshold, market research analysts at firms with fewer than 50 employees, technical writers, and first-year associates in commercial law.

That last one deserves attention. Law firms are not replacing associates with AI. They are not hiring as many of them. The workflow change is a reduction in headcount growth, not an elimination of existing jobs — but the distinction matters less to a 2025 law school graduate trying to find a first position than it does to an economist counting net employment levels.

The sectors where employment has held up against clear AI capability are more instructive than the sectors where it’s declining. Registered nurses: up 3.2% year-over-year. Elementary and middle school teachers: essentially flat, with regional variance explained by demographic factors. Construction managers and site supervisors: up 4.1%, driven partly by infrastructure spending but also by the genuine complexity of coordinating physical work that defies automation. Mental health counselors: up 6.8%, the highest growth in the professional services category, in a finding that startles people who haven’t thought about it but shouldn’t.

The mental health number is worth sitting with. AI-assisted therapy tools are widely deployed by 2026. BetterHelp, Woebot, and a dozen other platforms have AI triage and support layers. This has not suppressed demand for human counselors — it has, the data suggests, expanded it, by reducing the stigma barrier to help-seeking and by identifying cases that need human escalation. This is the complementarity story in its most literal form: the AI creates demand for the human. But note what kind of human it’s creating demand for. Not the junior counselor doing intake paperwork. The credentialed, experienced clinician conducting actual therapeutic work.

The geographic distribution of displacement is the least-covered and most important dimension of the current data. The employment contractions in AI-susceptible roles are not uniform across metropolitan areas. They’re heavily concentrated in mid-tier cities — the 50th to 150th largest metros in the US — where the local economy had a higher share of exactly the kinds of back-office, knowledge-work-adjacent service jobs that AI handles first. Cities like Dayton, Ohio; Bakersfield, California; Peoria, Illinois; and Columbia, South Carolina show employment contraction in office and administrative support occupations that is roughly three times the national average rate.

This is not a coincidence. It’s a structural consequence of the fact that mid-tier cities were the places where automatable knowledge work had clustered precisely because labor was cheaper there than in major metros. The business case for AI substitution is strongest where the human labor being substituted is cheapest. Which means the workers least able to absorb the transition are the ones experiencing it most acutely.

Goldman Sachs published a now-much-cited 2023 paper estimating that AI could affect 300 million full-time jobs globally. That paper’s methodology involved categorizing tasks within jobs as AI-exposable. By 2026, the criticism of that methodology is empirically grounded: task exposure is not the same as job displacement. A role where 40% of tasks are AI-exposable does not lose 40% of its employment. It typically either loses the role entirely (if those 40% of tasks were the justification for the role’s existence) or it retains the role at the same headcount while the productivity of the remaining 60% of tasks increases.

The actual displacement ratio — jobs lost relative to tasks exposed — appears to be running at roughly 15-25% in the first wave, which is lower than naive task-exposure models predicted but higher than the “AI just changes what workers do” optimists wanted to believe.

One number that has not gotten nearly enough attention: the ratio of AI-related job creation to AI-related job displacement. The World Economic Forum and OECD both track “AI-native roles” — positions that either didn’t exist before 2022 or have been substantially redefined by AI deployment. These include AI trainers, prompt engineers, model output auditors, AI deployment specialists, and a growing category of “AI liaison” roles at large organizations that coordinate between AI systems and the humans who work alongside them. The net creation in these roles through mid-2026 is real but modest. There are approximately 340,000 net new AI-native roles in the US labor market. The number of roles experiencing AI-driven contraction, measured by the BLS, is approximately 1.2 million.

The gap is not a disaster. It is not “fine.” It is a policy problem.

The thing that the data is beginning to make clear — which should have been clear from first principles but apparently wasn’t — is that job creation in response to technological displacement doesn’t happen automatically. It happened in previous technological transitions only because specific institutional conditions were present: government investment in new industries, education systems that retooled, geographic mobility supported by affordable housing, and (underappreciated) time. The industrial revolution created more jobs than it destroyed, but not on a timeline that helped the handloom weavers of 1820.

The 2026 data doesn’t settle the long-run debate about AI and net employment. The technology is still early, the complementarity effects are still developing, and the organizational learning curve for deploying AI effectively is still being climbed. What the data does settle is the near-term distributional question: who is absorbing the cost of the transition right now. The answer is what it almost always is in technological disruptions — the workers with the fewest options, in the places with the least slack, in the roles closest to the capability frontier of the new technology.

The honest read of the first-year data is that the disruption is real, the aggregate story is too early to write, and the specific stories of specific workers in specific places deserve more attention than the projections and counter-projections that have dominated the conversation so far.

Get the best articles from Think Different

One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.