Hiding in the mean: AI is adding jobs and removing them at the same time

Estimates of AI’s employment effect disagree in size and even in sign. This column argues the disagreement is built into the measurement: AI substitutes for humans on some tasks and complements humans on others, and any single exposure index nets these opposing forces into one number. Classifying US work tasks by what they require of a human and letting the two forces enter separately, both effects appear at once: since 2021, highly substitutable work has grown 3.3 percentage points a year more slowly, and complemented work 3.3 points faster. The adjustment runs primarily through hiring rather than firing, and it lands on the young.

In May 2025, Anthropic’s chief executive warned that AI could eliminate half of entry-level white-collar jobs within five years (Axios 2025). Sixteen months later, US unemployment stands at 4.1%. Nvidia’s chief executive counters that AI creates jobs too: “the number of radiologists has gone up” since it took over reading scans (NVIDIA 2026). Economists disagree just as sharply: MIT’s David Autor rejects a “software apocalypse”, Virginia’s Anton Korinek expects labour’s income share to fall by half, and Yale’s Martha Gimbel splits the difference (Anders 2026). The Economist (2026) concluded that “the jobs apocalypse is postponed”, counting roughly a million US jobs created against some 200,000 lost. The same week, the New York Times reported the opposite mood: slower hiring and weaker wage growth in AI-exposed occupations (Smith 2026), drawing partly on the evidence below.

Each camp has data. The dispute does not settle – almost everyone in it is measuring the same wrong object.

Tasks, not jobs

AI does not act on jobs. It acts on tasks, and a job is a bundle of heterogeneous ones. Ask how much AI speeds up a lawyer and the question has no answer: AI drafts a brief in minutes but does nothing for a court appearance. This is the canonical unit of the technology-and-labour literature (Autor et al. 2003, Acemoglu and Restrepo 2019). Yet nearly every published AI exposure measure is built at the occupation level, one number per job, because grading 17,536 task statements by hand was never feasible. Large language models have removed that constraint, but most of the field has not used them to fix it.

The problem with one number per job runs deeper than imprecision: AI can push the same job in two opposite directions at once. On some tasks it substitutes for the human, and employment in that task falls. On others it complements the human: it does part of the work, cuts the cost of the output, and raises demand for the human who finishes it.

The consequence is visible in one of the gold standard measures, the AI Occupational Exposure (AIOE) index of Felten et al. (2021), built before language models made task-level grading feasible. It scores medical secretaries and secondary school teachers just 0.027 apart on a scale spanning several points – effectively identical exposure. Yet AI can already handle much of a secretary’s job end to end (drafting summaries, transcription, scheduling) while it cannot deliver the part of a teacher’s job that matters most: standing in front of a classroom. Two thirds of the secretaries’ tasks are fully substitutable, against roughly a fifth of the teachers’, while the teachers’ work holds five times the complement share. A single occupation-level index cannot express that: one number for two jobs pulled by opposite forces of very different strength. Between them, these two occupations employ nearly two million Americans.

Measuring both effects of AI

In Verschuere and Cameron (2026a), we grade every task statement in the US occupational database by one structural question: what does this task require of a human? Three answers are possible: (1) nothing, i.e. the output can be produced end to end by AI, such as drafting a contract or filling out a form (substitution, S); (2) a person physically present in real time, or a licensed sign-off, such as a therapy session or live classroom teaching (complement, C); or (3) physical action AI cannot perform, such as plumbing, construction, or cutting hair (inert, I). Every occupation then carries three shares instead of one net score.

Other researchers are independently converging on the same substitute/complement split from different data: a time-use study built its own such measure and found the two forces pulling weekly working hours in opposite directions (Jiang et al. 2025).

The approach also sidesteps a problem that dogs newer LLM-graded task-level scores such as Eloundou et al. (2024), previously the field’s default: scored with different models, the same task data gives different answers. The share of US occupations rated highly exposed ranges from 2.7% under one model to 51.5% under another, a nearly twentyfold spread on identical data (Yin 2026). Our criterion avoids this by asking what a task structurally requires rather than what a model can currently do: six models from four developers agree unanimously on 82% of tasks (Verschuere and Cameron 2026b).

Applied to the full US occupational database, employment-weighted work comes out 32% substitution, 15% complement, and 53% inert (Figure 1), roughly a two-to-one ratio. That split is itself a finding: only a third of the economy’s work is something AI could do end to end today, and the complement share – work AI makes more valuable rather than replaces – is less than half that size. The remaining majority sits in tasks AI cannot touch: the physical economy, from plumbing to construction. That two-to-one gap is noteworthy: it marks the two pools of work AI can act on, and shows there is twice as much of the former as the latter.

Figure 1 Every US occupation mapped by its substitution and complement shares

AI adds jobs and takes them away

Taken to occupation-level employment data since 2021, both forces show up in the same difference-in-differences regression, with opposite signs. Work AI can substitute has grown 3.3 percentage points a year more slowly than work AI does not touch; work AI complements has grown 3.3 points a year faster (Figure 2). The complement estimate is significant under both conventional and clustered standard errors; the substitution estimate is significant under the former, only marginal under the latter. Neither margin predicts employment growth in the 2019–21 window, before AI could plausibly matter, and the same test run entirely inside 2015–2019, years before ChatGPT existed, finds nothing either, both checks statistically insignificant. The result also survives controlling for the COVID rebound and the rate cycle directly. 

Figure 2 The scissors: Relative employment paths of substitutable and complemented work, 2015–2025

In headcount terms, this is roughly 1.5 million jobs a year removed through substitution against 0.7 million added through complementarity. The two gradients run at the same speed, so the net comes from composition. This also explains the literature’s disagreement mechanically: alongside the two-margin composition, published single indices, including our own aggregate scalar, lose statistical significance, each inheriting a different mix of the two opposing forces. The aggregate looks calm only because it nets two opposing flows.

This same granularity helps understand how the adjustment happens. Inside the most substitutable occupations, separation rates did not rise after ChatGPT, so the change is not layoffs. It shows up instead in hiring, and it lands on the youngest: job starts for workers aged 22 to 25 in the most substitutable quarter fell by a third, roughly 227,000 entry positions a year that no longer open, with no matching decline in young hiring into work AI cannot substitute. Payroll, resume, and posting data show the same asymmetry elsewhere (Brynjolfsson et al. 2025, Hosseini Maasoum and Lichtinger 2025).

The impact is also morphing over time: within substitutable work, the effect is climbing the pay scale, significant first in low-paid, high-turnover occupations, now significant in middle-paid ones too.

Policy implication

As shown here, AI’s effect on jobs runs in both directions at once, removing substitutable work and adding complementary work, and different groups are affected differently. So far, the impact falls hardest on young entrants into the most substitutable occupations, while incumbents and complement-heavy occupations are largely untouched. That pattern is not static either; the composition of the economy, and its effect on employment, will keep moving as the technology and its adoption evolve.

Tracking the nuances of AI labour impact requires looking at both margins together, not one blended number. Better tracking of where and how AI operates is what makes better understanding, and better policymaking, possible.

AI is a genuinely transformative technology, and this column has shown it is already reshaping the economy and society, in both directions at once. This modelling offers not a verdict on that reshaping but the map – a way to keep tracking its evolution.

Source : VOXeu