For three years the AI-jobs debate has produced a different number every month. A new analysis argues almost everyone in it is measuring the wrong object — because AI adds jobs and removes them at the same time, and any single “AI exposure” score nets the two opposing forces into one misleading figure.
🔍 THE BOTTOM LINE: Graded task by task, AI-substitutable work in the US has grown about 3.3 percentage points a year slower since 2021, while AI-complemented work has grown 3.3 points a year faster. The adjustment is running through hiring rather than firing — and it is landing hardest on workers aged 22 to 25.
One number per job was the mistake
Researchers Bruno Verschuere and Andrew Cameron grade every task statement in the US occupational database with a structural question: what does this task require of a human? Three answers — nothing (AI can do it end to end), a physically present person or licensed sign-off (a therapy session, a live classroom), or physical action AI cannot perform (plumbing, construction, cutting hair).
Applied across the database, employment-weighted work comes out roughly 32 percent substitution, 15 percent complement, and 53 percent inert (Verschuere & Cameron 2026) — a two-to-one split the authors call a finding in itself: only about a third of the economy’s work is something AI could do end to end today, and the share AI makes more valuable is less than half that. The remaining majority sits in the physical economy, untouched.
The task method exposes why older, single-score studies disagree so wildly. The widely used Felten exposure index rates medical secretaries and secondary school teachers just 0.027 apart — effectively identical. Yet about two-thirds of a medical secretary’s tasks are fully substitutable against roughly a fifth of a teacher’s, while teaching holds five times the complement share. Two occupations employing nearly two million Americans were being scored as the same job; a review of newer LLM-scored indexes found the share of “highly exposed” occupations ranges from 2.7 percent to 51.5 percent depending on which model did the scoring.
The scissors: two forces, same speed
Since 2021, work AI can substitute has grown about 3.3 percentage points a year slower than untouched work; work AI complements has grown about 3.3 points a year faster. In headcount terms the authors estimate roughly 1.5 million jobs a year removed through substitution against 0.7 million added through complementarity — the net looks calm in aggregate because two large flows cancel.
The checks matter. Neither effect predicts employment growth in 2015–2019, before ChatGPT existed; the same test finds nothing in the pre-AI window, and the result survives controlling for the COVID rebound. The complement estimate holds under both conventional and clustered standard errors; the substitution estimate is significant under conventional tests and only marginal under clustered ones — the authors flag that honestly.
It lands on hiring, not firing — and on the young
The part with the most direct relevance to anyone under 26: inside the most substitutable occupations, separation rates did not rise after ChatGPT. Nobody is being fired out. Job starts for workers aged 22 to 25 in the most substitutable quarter fell by about a third — roughly 227,000 entry positions a year that no longer open. The door did not close; it stopped opening.
The squeeze is also climbing the pay scale: first significant in low-paid, high-turnover occupations, now significant in middle-paid ones too. And the growth side is real — complement-heavy work, where AI handles part of the task and raises the value of the human who finishes it, has been expanding at the same 3.3-point pace throughout. That dovetails with the Stanford “canaries” finding that entry-level employment in AI-exposed roles has fallen while experienced roles hold, and with the entry-level hiring freeze documented earlier this year.
The apocalypses that aged sideways
The study’s framing is a quiet correction to two loud predictions. In May 2025 Anthropic’s chief executive warned AI could eliminate half of entry-level white-collar jobs within five years; sixteen months later US unemployment sat at 4.1 percent. In the other direction, Jensen Huang’s “the number of radiologists has gone up” and The Economist’s September verdict — roughly a million US jobs created against some 200,000 lost, “the jobs apocalypse is postponed” — both capture a real half of the same picture. The New York Times’ reporting on slower hiring and weaker wage growth in exposed occupations was capturing the other half. Everyone, on this reading, was measuring one blade of the scissors.
What it means for New Zealand
New Zealand’s labour market is smaller and slower-moving, but the mechanism travels: the Reserve Bank’s analysis of AI and the graduate first rung flagged the same entry-point vulnerability, and local job-ad data shows AI skills growing as a share of postings while office and admin listings flatline — the complement and substitution blades visible in one national dataset. For young Kiwis the study reframes the advice: incumbency is currently protective (no layoff wave), but entering a substitutable occupation is harder every year, while fields where AI assists a licensed or physical human — teaching, health, trades, senior technical oversight — are the ones expanding. That matches what this site has seen in New Zealand’s own application-flood data: the robots screen entry, the humans still hire judgment.
❓ FAQ
What does “two margins” mean in this study? The researchers separate AI’s effect on jobs into substitution (AI does the task end to end) and complementarity (AI does part of it and makes the human’s remaining work more valuable). Older studies blended both into one “exposure” score, which is why their conclusions disagreed.
Is AI causing mass layoffs in the US? Not on this evidence. Separation rates in the most substitutable occupations did not rise after ChatGPT; the adjustment shows up in hiring, with about 227,000 fewer entry-level job starts a year for workers aged 22 to 25, per the study’s estimates.
Which jobs grew and which shrank? Work AI can substitute — drafting, form-filling, document processing — grew about 3.3 points a year slower since 2021. Work AI complements — live teaching, therapy, licensed sign-offs, physical trades — grew about 3.3 points a year faster.
Is this peer-reviewed? No. It is a working paper (SSRN 7195359) published as a CEPR/VoxEU column on 2 October 2026, with its own uncertainty flags — the substitution effect is only marginal under clustered standard errors. Treat it as a promising method, not a settled verdict.
🔍 THE BOTTOM LINE
The study’s map is more useful than its verdict: AI is not a bulldozer moving one direction across the labour market, it is a scissors opening in both directions at once — and the pivot point is whoever is trying to get in. For workers already inside, the firing wave keeps failing to arrive; for new entrants, the maths says pick the blade that is rising, not the one that is falling.
📰 Sources
- CEPR/VoxEU — Hiding in the mean: AI is adding jobs and removing them at the same time (2 October 2026)
- Verschuere & Cameron — “Hiding in the mean: The two margins of AI’s employment effect” (SSRN working paper 7195359)
- CEPR/VoxEU — When the ruler is made of the thing it measures (12 May 2026)
- The Economist — “The jobs apocalypse is postponed. An AI jobs boom is here” (4 September 2026)
- The New York Times — “The quiet way A.I. is hitting the work force” (16 September 2026)
- Axios — interview with Dario Amodei on entry-level displacement (28 May 2025)
- Stanford Digital Economy Lab — Brynjolfsson, Chandar & Chen, “Canaries in the coal mine?” (2025)
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
