Which tasks and occupations will be automated, augmented, transformed, or newly created — and how quickly?
Exposure studies flag most occupations as partially exposed; real usage data shows adoption concentrated in software, writing, and analysis. The augmentation-vs-automation split within tasks is the number to watch.
View on the map → · Open in Browse →
The Remote Labor Index's measured automation rate roughly doubled in one model generation — a capability fact, not yet a labor-market one.Evidence: 1
End-to-end automation of real projects is improving fast on direct measures — a fact about what models can do, not yet about employment.
Fable 5 reaches the highest automation rate measured so far, 15.8%, roughly double Opus 4.8 at 8.3%.
Remote Labor Index update: the measured automation rate of real freelance projects roughly doubled in one model generation.
Exposure and adoption predict restructured work and shifted hiring, not employment loss — so far.
task-based measures of technological exposure are better understood as indicators of occupational restructuring than as direct forecasts of employment decline
New occupational-outcomes evidence for 2019–24: AI exposure predicts restructuring of work rather than employment decline.
people who use Claude in the most automated way expect AI to take on more of their tasks in the next year, yet feel the most optimistic about what that means for their work
Only ~10% of workers rate their own job loss likely within a year; over a third fear for junior colleagues.
growing slower post-ChatGPT compared to a no-AI counterfactual, by about 3 percentage points per year
AI compresses coding but not accountability; AI-attributed layoffs are largely narrative cover for cost-cutting.
no relative deterioration in unemployment among workers whose jobs contain a large share of tasks that Claude is used to automate
Anthropic's head of economics, on Anthropic's own usage data through June 2026 — a finding that cuts against his CEO's displacement forecast.
What the benchmarks, exposure measures, and usage data capture — and what they cannot.
High scores on the benchmarks, therefore, would not imply end-to-end automation of digital professions.
Methodological critique: GDPval and RLI tasks are too self-contained to forecast full occupational automation.
Proposals and positions from institutions, included for what they reveal as much as what they argue.
An institutional position as much as an argument: a lab CEO proposes tiered economic responses — wage insurance up to income support — pegged to measured AI-driven displacement.
Pilot brief — draft v5 · editorial review pending. Core article last revised August 2026; evidence updated August 2026.
Written by Drafted with AI assistance (Claude), revised through editorial feedback · edited by Elliott Lehrer
Scope: Primarily cognitive work in advanced economies, mid-2020s through the 2030s. Embodied automation, developing-country labor markets, and the long-run post-AGI case are treated where they bear on that question, not surveyed comprehensively.
Which kinds of work will AI automate outright, which will it augment, which will it transform into something new, and which will it leave largely untouched — and on what timetable? Automation, augmentation, transformation, and creation describe what happens to tasks. Complementarity and substitution describe what those changes ultimately do to the demand for human labor. This entry answers both: the task-and-occupation map, and the deeper question that determines what the map means.
When an AI system enters a workplace, it can do two things to the humans there. It can complement them — raise the value of their time by handling sub-tasks, extending their expertise, or letting them serve more people — or it can substitute for them, performing their work well enough that the human's contribution commands less pay, or none. The question is which force dominates, for most workers, over the horizon in which this generation of institutions must respond.
Three distinctions keep this question answerable, and conflating them is a pervasive error in the public debate:
Three different questions hiding in one word
- Usage: Is the AI operated by a worker during their work, or running instead of one?
- Productivity: Does it raise output per worker-hour?
- Labor demand: Do firms end up wanting more or fewer workers, at higher or lower pay?
Only the third is economic complementarity in the sense that matters for employment and wages. A tool can double each worker's productivity while letting a firm halve its headcount; automating a task outright can increase employment if it cuts prices enough to expand demand. Augmentative use and productivity gains are evidence about mechanisms — they do not settle labor demand.
Two further clarifications. The unit of analysis is the task, not the job: almost no occupation is a single task, so "AI can do X% of tasks in occupation Y" tells you little by itself. Removing tasks from a job can make the remaining human tasks more valuable — ATMs sharply reduced the tellers needed per branch, but cheaper branches led banks to open more of them, which for years offset the expected decline in teller employment (Bessen 2016). And complement-versus-substitute is not a fixed property of the technology: it is a moving ratio, set by where the capability frontier sits, how fast firms actually reorganize around it, and — on one influential view — deliberate choices about what gets built. A system that complements a radiologist in 2026 may substitute for one in 2030; a system that substitutes for a junior analyst may complement the senior one reviewing its output.
This question is distinct from its neighbors: the size of AI's aggregate growth effects (3.1.1) can be answered independently of who captures them; the wage-bargaining channel (3.2.2) operates even without displacement; the entry-level question (3.2.3) concerns the sequence of substitution; and who owns the machines (3.5.1) matters most in exactly the scenarios where substitution wins.
If AI primarily complements labor, its gains flow substantially through wages — wage-mediated distribution remains a plausible central channel, and policy can focus on adoption, training, and adjustment. If AI primarily substitutes for labor, gains flow to whoever owns compute, models, and capital; wages can fall even while output soars; and distribution comes to depend much more heavily on ownership, taxation, and transfer mechanisms that democracies have historically found harder to enact than employment policy. Which world we are entering determines how broadly "learn to work with AI" remains useful advice, and whether the right policy portfolio is retraining programs or a renegotiation of who owns the productive base. Note that the comfortable case is not automatically comfortable: complementarity can coexist with weakened bargaining power, rents concentrating with model providers, and a falling labor share — which is part of why the wage question (3.2.2) stays open even if this one resolves toward complement.
What changes first: Bounded, digitally mediated, readily evaluated cognitive tasks — mapped in Which work changes first? below.
What the evidence shows: Large productivity gains on some tasks, with the largest gains to the least experienced; real but narrow substitution in standardized-text markets; and no identifiable aggregate employment or wage effect yet.
Where the dispute concentrates: Young and entry-level workers in highly exposed occupations — a signal that is real in some datasets and contested in its cause.
Why confidence remains limited: Capability benchmarks are improving much faster than firms and labor markets reorganize, which makes extrapolation beyond a year or two unreliable in both directions.
The complement-substitute lens explains why task automation does not translate mechanically into job loss. It does not, by itself, say which work changes first. The best current predictors are task characteristics, not occupation labels:
| Task characteristic | Likely near-term effect |
|---|---|
| Digital output, clear specification, cheap evaluation | Earliest automation |
| Structured work with a skilled human checking outputs | Augmentation, then partial automation |
| Heavy context-switching and tacit organizational knowledge | Slower, more uneven transformation |
| High legal, reputational, or safety accountability | Capability may long precede deployment |
| Physical dexterity in unstructured environments | Gated on robotics progress, not language models |
| Output whose demand expands as costs fall | Automation can raise employment |
| Trust, negotiation, or human-preference work | Rebundling more likely than replacement |
Mapped onto the current landscape, using the evidence cited below: software development is furthest along the augmentation path — the deepest adoption in usage data, with agents beginning to absorb bounded implementation work while design, review, and integration stay human. Translation and routine commercial writing are the substitution frontier: digital output, forgiving quality tolerance, thin client relationships. Customer support and structured analysis are mid-transition, showing the novice-uplift pattern — the assistant carries the bottom of the skill distribution. Accountability-bearing professions (law, medicine, finance) show capability running well ahead of deployment: models pass the benchmarks while liability, regulation, and trust gate the jobs. Exposed office occupations are so far reorganizing rather than shrinking — new AI-related tasks and shifted duties without net job loss, the pattern documented in Denmark's administrative data. Manual and embodied work remains gated on robotics rather than language models, which is why this entry's scope excludes it — but the boundary is where most global employment lives.
On newly created work: so far, new tasks are appearing mostly within existing occupations — AI oversight, evaluation, prompt and agent orchestration, output verification — rather than as new occupational categories. That is consistent with the 1940–2018 record, in which new work typically began as new specialties inside old jobs before becoming visible in official classifications (Autor, Chin, Salomons & Seegmiller 2024).
On timing: change concentrates where three things align — sufficient capability, cheap verification of output, and low accountability for error. Work protected by any one of the three has, so far, seen its timetable slip from years of predicted disruption to gradual workflow change.
Four serious positions, plus one that cuts across them.
1. Complementarity, as usual (the augmentation view). Technology has always destroyed tasks and created more valuable ones; AI is following the pattern. The historical anchor: roughly 60% of current US employment is in job specialties introduced since 1940 (Autor, Chin, Salomons & Seegmiller 2024). Autor's forward-looking version is specific to AI: labor markets pay for expertise, and AI can extend it — letting nurse practitioners do work once reserved for doctors, competent generalists work once reserved for credentialed specialists — thereby rebuilding the middle-skill jobs computerization hollowed out (Autor 2024; Autor & Thompson 2025). Bessen's demand-elasticity mechanism supplies the economics: automation raises employment where cheaper output meets expandable demand (Bessen 2018).
2. Complement now, substitute at the frontier (the capability-contingent view). Today's complementarity is real but transitional — an artifact of where the capability frontier happens to sit. Korinek and Suh's scenario modeling formalizes the endgame: as automation approaches completeness, wages first rise, then collapse, because the scarce factor shifts from human time to compute (Korinek & Suh 2024); Epoch's GATE model reaches similar conclusions across a wide parameter range (Epoch AI 2025). Susskind's "task encroachment" makes the gradualist version: machines need not replicate human cognition to keep absorbing tasks, and the complementing forces that protected labor through past transitions weaken as encroachment proceeds (Susskind 2020). On this view, the right question is not whether substitution but when.
3. Substitution is already arriving (the displacement view). The mechanism: frontier systems are improving rapidly on real economic deliverables, agents are beginning to complete end-to-end work rather than sub-tasks, and the first displacement signals — young workers in exposed occupations, freelance text markets — are appearing exactly where theory predicts substitution would show first. What looks like aggregate calm is, on this reading, deployment lagging capability, temporarily. A prominent strong-form forecast comes from Dario Amodei: half of entry-level white-collar jobs eliminated within one to five years, with unemployment reaching 10–20% (Axios interview, 2025) — an accurately quoted claim worth weighing alongside the fact that its author leads a frontier lab.
4. Diffusion dominates capability in the near term (the normal-technology view). The binding constraint is not what models can do but how fast institutions absorb them: like electricity and computing, AI will take decades to reorganize work, because adoption requires process redesign, trust, complementary investment, and organizational change (Narayanan & Kapoor 2025). Acemoglu's task-based arithmetic reaches a compatible conclusion from the cost side: modest aggregate effects (~0.7% TFP over a decade) under current capabilities (Acemoglu 2024). This camp does not deny productivity complementarity; it denies that either complementarity or substitution will transform aggregates soon. Its trump cards are the null results below.
Cutting across all four: substitution as a choice. Brynjolfsson's "Turing Trap" argument holds that complement-versus-substitute is partly a design variable: building AI to imitate humans biases it toward substitution, building it to extend humans biases it toward complementarity, and current incentives — including tax codes that price labor above capital — tilt toward imitation (Brynjolfsson 2022). Acemoglu, Autor and Johnson's pro-worker-AI agenda is the policy expression (2026). This position changes what the question is — from a forecast to a decision.
The disagreement persists for identifiable reasons, not lack of data. The camps extrapolate from different curves: position 3 extrapolates capability benchmarks, which are improving very fast; position 4 extrapolates adoption and macro data, which are barely moving; both describe their own curve accurately. They measure different objects — tasks, jobs, wages, and labor share can move in different directions simultaneously. They also reason at different equilibrium levels: a firm may need fewer workers after automating a task, while its industry employs more because lower costs expand demand, while the economy's employment depends additionally on spending, new industries, and monetary and fiscal response — partial-equilibrium intuitions and general-equilibrium outcomes regularly point in opposite directions, and much public argument switches between them unannounced.
The deepest conceptual uncertainty is reinstatement (Acemoglu & Restrepo 2019): the historical record shows new tasks absorbing displaced labor, but new-task creation is precisely what no one knows how to forecast, and the historical base rate is informative only if AI is a normal technology — which is itself the contested question. Horizon ambiguity does quiet work too: many apparent disagreements dissolve into "complement for the next five years, unknown after ten," consistent with almost everyone's model and almost no one's rhetoric. Finally, there is an unavoidable prior about whether human cognition retains durable comparative advantage anywhere once machines reason cheaply — an assumption current evidence cannot discipline.
Most early task-level experiments show substantial productivity gains and compression of performance gaps on bounded tasks. ChatGPT cut professional writing time ~40% and raised quality ~18%, with the weakest writers gaining most (Noy & Zhang 2023). A generative assistant raised call-center issues-resolved-per-hour ~15% on average but ~30%+ for novices, with minimal effects on the most experienced (Brynjolfsson, Li & Raymond 2025). Copilot sped programmers ~56% on one well-scoped implementation task (Peng et al. 2023). The BCG field experiment found consultants ~25% faster with 40%+ higher quality inside AI's "jagged frontier" — and worse than controls outside it (Dell'Acqua et al. 2023). These results establish operational assistance and skill compression. They do not establish whether firms ultimately employ more or fewer people — none measured equilibrium labor demand.
The flagship counter-result is a warning about measurement, in both directions. METR's randomized trial found experienced open-source developers took ~19% longer with early-2025 AI tools — while believing they had been sped up ~20%, a roughly 39-point perception gap (METR 2025). A 2026 follow-up produced point estimates suggesting modest speedups (~18% for returning participants, ~4% for new ones, confidence intervals spanning zero), but METR judged the estimates unreliable: developers unwilling to work without AI increasingly selected out of the study, likely biasing results against AI, and the organization concluded the design has become a poor proxy as AI use normalizes (METR 2026). The durable lessons: self-reported productivity is untrustworthy, and clean measurement is getting harder precisely as the answer matters more.
Aggregate data remain quiet — and the best micro data show reorganization without displacement. The aggregate claim rests on economy-wide tracking: Yale's Budget Lab, following mid-2026 CPS data with a synthetic difference-in-differences design, finds US occupational churn near historical norms and no clear relationship between AI exposure and employment outcomes (Budget Lab 2026). Beneath the aggregates, the sharpest adopter-level result is a micro null: Denmark's linkage of 25,000 workers in exposed occupations to administrative records finds AI adoption produced effects on adopters' earnings and hours statistically indistinguishable from zero — confidence intervals ruling out effects above ~2% two years in — while documenting substantial reorganization of work, new AI-related tasks, and some movement into higher-paying, more AI-relevant roles (Humlum & Vestergaard 2025, rev. 2026 as Still Waters, Rapid Currents). That combination — structural change before aggregate change — may be the most informative single result in the literature, but it is evidence about adopting workers, not about equilibrium labor demand.
The entry-level margin is where the fight is. Stanford's "Canaries in the Coal Mine" found a ~13% relative employment decline for 22–25-year-olds in the most AI-exposed occupations — revised toward ~16% through October 2025 under broader controls — concentrated where AI automates rather than augments, with the clean divergence beginning in 2024 (Brynjolfsson, Chandar & Chen 2025). The counter-case is substantial: postings in exposed occupations peaked months before ChatGPT and track the Fed's tightening cycle (Iscenko & Curto Millet 2026); New York Fed researchers attribute roughly two-thirds of the recent rise in young-graduate unemployment to remote work destroying mentorship rather than to AI (Federal Reserve Bank of New York 2026); and Anthropic's matched analysis finds no general unemployment increase for exposed workers but a marginally significant ~14% decline in job-finding rates for the youngest entrants (Anthropic 2026). Recent-graduate unemployment (~5.7%) remains above the general rate — historically unusual — for genuinely contested reasons (NY Fed 2026). This entry treats the dispute as unresolved.
Capability is running well ahead of deployment. On GDPval — real deliverables from 44 occupations, graded blind against experts averaging 14 years' experience — the best model went from far below parity in 2024 to 47.6% win-or-tie in September 2025, and OpenAI's December 2025 release reported 70.9% for GPT-5.2, versus 38.8% for its predecessor months earlier (OpenAI 2025a; OpenAI 2025b) — the latter a self-reported score from the developer's launch materials, on the developer's own benchmark, and worth the same conflict-of-interest discount this entry applies to lab forecasts elsewhere. The benchmark's limits matter as much as its trend: tasks are well-specified, one-shot deliverables — no task discovery, no iterative client interaction, no organizational integration, and no measurement of whether using the output changes headcount. On the harder Remote Labor Index — complete freelance projects, end-to-end — the best system fully automated 2.5% of projects at professional quality at launch in October 2025 and 15.8% by July 2026, a more-than-sixfold rise in under nine months, with most outputs still failing professional acceptance (CAIS & Scale AI 2026). Meanwhile Anthropic's usage data consistently show API deployments running more directive, automation-style workflows than conversational use, with the precise shares shifting between reporting periods (Anthropic Economic Index 2026) — consistent with substitution, where it comes, arriving through deployed pipelines rather than chat windows.
Where displacement is already documented, it is narrow and specific. Freelancers in highly AI-exposed categories on Upwork saw monthly earnings fall ~5.2% and jobs ~2% after ChatGPT (Hui, Reshef & Zhou 2023). Machine-translation adoption — beginning well before generative AI — foreclosed an estimated ~28,000 translator positions that would otherwise have been created in higher-adoption US markets, 2010–2023 (Frey & Llanos-Paredes 2025); 36% of surveyed UK translators report work already lost to generative AI (Society of Authors 2024). These are real substitutions — concentrated in markets where a large share of the paid output is standardized text, quality tolerance is forgiving, and client relationships are thin.
Discriminating indicators, grouped by what they measure:
Capability. GDPval-style expert parity and Remote Labor Index end-to-end automation continuing their current trajectories — RLI moving from ~16% through 40–50% — versus stalling at the messy-work frontier of task discovery, iteration, and integration. This tests the technical frontier alone, separate from whether anyone deploys it.
Adoption. Whether benchmark capability translates into paying deployment: agent workflows in production, automation-oriented API usage growing rather than plateauing, firms redesigning processes rather than layering AI onto existing ones. Capability without commensurate adoption strengthens the diffusion-first account; adoption that closely tracks capability would weaken it.
Labor demand. The entry-level identification fight resolving — the Stanford signal surviving the interest-rate and remote-work critiques through 2026–27 data and spreading up the age curve, or dissolving into macro confounders. Postings, hiring rates, and hours in exposed occupations breaking from rate-cycle explanations.
Distribution. Labor share of income — the variable positions 2 and 3 ultimately predict will fall, and which has not yet moved for demonstrably AI-related reasons. Also: wage premia emerging for AI-direction, judgment, and taste skills, which complementarity predicts.
Reinstatement. The forecastable version of the unforecastable — watched at the task level, not just in census categories: new tasks appearing within existing occupations, changed job descriptions, newly demanded skills, new role titles, and only eventually new occupational classifications. What such work pays is a strong test of position 1's historical mechanism operating in the AI transition.
If complementarity dominates, the policy portfolio is familiar and tractable: diffusion support, training, credential reform, and competition policy to keep gains from concentrating. If substitution dominates, wage-mediated distribution weakens structurally, and the load shifts toward ownership and transfer mechanisms (3.5.3) under political conditions — falling labor power — that make them hardest to enact; the entry-level sequence (3.2.3) additionally strains the apprenticeship pipeline through which expertise reproduces (3.2.4). If diffusion dominates capability for years, the main near-term risks are premature policy and misallocated attention. And if the design-choice view is right, part of the answer is endogenous to decisions — tax treatment of labor versus capital, procurement, evaluation norms — being made now, mostly without being recognized as decisions.
Best introduction: Autor, AI Could Actually Help Rebuild the Middle Class (Noema, 2024).
The framework: Acemoglu & Restrepo, Automation and New Tasks (JEP, 2019); Autor, Chin, Salomons & Seegmiller, New Frontiers (QJE, 2024); Autor & Thompson, Expertise (NBER, 2025).
The case for eventual substitution: Korinek & Suh, Scenarios for the Transition to AGI (NBER, 2024); Epoch AI, GATE (2025); Susskind, A World Without Work (2020).
The diffusion-skeptical case: Narayanan & Kapoor, AI as Normal Technology (2025); Acemoglu, The Simple Macroeconomics of AI (2024).
Substitution as a design choice: Brynjolfsson, The Turing Trap (Daedalus, 2022); Acemoglu, Autor & Johnson, Building Pro-Worker Artificial Intelligence (NBER, 2026).
Ongoing trackers: Anthropic Economic Index · Remote Labor Index · GDPval · NY Fed recent-graduate labor market · Yale Budget Lab tracking
Every source cited in this entry. Acemoglu (The Simple Macroeconomics of AI, 2024) · Acemoglu & Restrepo (Automation and New Tasks, 2019) · Acemoglu, Autor & Johnson (Building Pro-Worker Artificial Intelligence, 2026) · Amodei (Axios interview, 2025) · Anthropic (Labor market impacts of AI, 2026) · Anthropic (Economic Index; January 2026 report) · Autor (AI Could Actually Help Rebuild the Middle Class, 2024) · Autor, Chin, Salomons & Seegmiller (New Frontiers, 2024) · Autor & Thompson (Expertise, 2025) · Bessen (How Computer Automation Affects Occupations, 2016; AI and Jobs: The Role of Demand, 2018) · Brynjolfsson (The Turing Trap, 2022) · Brynjolfsson, Chandar & Chen (Canaries in the Coal Mine, 2025) · Brynjolfsson, Li & Raymond (Generative AI at Work, 2025) · The Budget Lab at Yale (Tracking the Impact of AI on the Labor Market, 2026) · CAIS & Scale AI (Remote Labor Index, 2026) · Dell'Acqua et al. (Navigating the Jagged Technological Frontier, 2023) · Epoch AI (GATE, 2025) · Federal Reserve Bank of New York (Remote Work Leaves Younger Workers Sidelined, 2026; The Labor Market for Recent College Graduates) · Frey & Llanos-Paredes (Lost in Translation, 2025) · Hui, Reshef & Zhou (The Short-Term Effects of Generative AI on Online Labor Markets, 2023) · Humlum & Vestergaard (Still Waters, Rapid Currents, 2025, rev. 2026) · Iscenko & Curto Millet (Looking for the Ladder, 2026) · Korinek & Suh (Scenarios for the Transition to AGI, 2024) · METR (developer RCT, 2025; update, 2026) · Narayanan & Kapoor (AI as Normal Technology, 2025) · Noy & Zhang (Experimental Evidence on the Productivity Effects of Generative AI, 2023) · OpenAI (GDPval, 2025; Introducing GPT-5.2, 2025) · Peng et al. (The Impact of AI on Developer Productivity, 2023) · Society of Authors (survey, 2024) · Susskind (A World Without Work, 2020)
Cite: “Which tasks and occupations will be automated, augmented, transformed, or newly created — and how quickly?” The Biggest Questions About AI, Elliott Lehrer (ed.), August 2026. Pilot brief — draft v5.