Anton Korinek

Who he is

Affiliation: Professor of Economics, University of Virginia (Economics & Darden), and Faculty Director of UVA's Economics of Transformative AI Initiative (EconTAI, launched 2025-08/09) — currently on leave, serving as Head of Transformative AI Economic Studies at Anthropic. Role check run 2026-09-21: stated by Darden News (2026-07-15), confirmed by his authorship of the Anthropic Institute's September-2026 scenarios paper. No start date or announcement found; the window is 2026-02-02 (the PIIE release lists UVA only) to 2026-07-15. Also: nonresident senior fellow (an affiliated-scholar seat), Peterson Institute (2026-02-02); nonresident senior fellow, Brookings; Anthropic Economic Advisory Council (2025-04-28); NBER; CEPR AI Research Network lead; Stanford Digital Economy Lab fellow; G7 High-Level Panel of Experts on AI (report 2024-12-19); TIME100 AI (2025-08). Source conflict: his own site, the Darden directory and Wikipedia still show no leave; the dated institutional item plus the lab's author list win over bio pages, which lag by months.
One-line position: If AI becomes truly transformative, the economic consequences — for growth, wages, labor share (the slice of national income paid out as wages) and inequality — are far more varied and extreme than standard economic models predict, and we need formal theory to navigate the range of possibilities. Since 2026 he says it from inside a frontier lab, with 16 Nobel laureates co-signing the urgency.

Discipline & technical bet

Macro and growth economist (Columbia PhD; earlier work in international finance) who treats "machines can do all economically valuable work" as a scenario to model, not a premise to mock. Toolkit: task-based growth models where work is a ladder of tasks and automation climbs it (Korinek–Suh, 2024); growth theory for fully automated production and research (Trammell–Korinek, 2026; Davidson–Halperin–Houlden–Korinek, 2026); and, new this cycle, a calibrated 2026–2030 simulation built at Anthropic that turns two dials — what share of cognitive tasks AI can do, and how fast firms adopt it — into paths for GDP, labor share, wages and unemployment (Anthropic Institute, 2026-09). The wager is methodological: those two dials are the right state variables, and data can eventually pin them down. He is building the weather model for the storm, not forecasting whether it comes — and as of 2026 the weather station sits inside the company whose product is the weather.

Key claims (Says)

• Empirical: Labor share could collapse under transformative AI — roughly two-thirds of all income and more than two-thirds of tax revenue derive from labor; if AI substitutes for most human work, this collapses. Held — still endorsed verbatim (Richmond Fed, Q4 2025) and now quantified: his model's extreme branch takes labor share from ~60% to 45.2% by 2030, the modest branch only to 59.4% (Anthropic Institute, 2026-09). Market-level check: BLS nonfarm labor share hit 52.8% in Q2 2026, the lowest since 1947 (BLS, 2026-09-15) — direction consistent, but the slide dates from 2000 (63.9%), predates the AI boom, and his own modest scenario predicts almost no AI effect through 2030. Held as a conditional.
• Empirical: Wide range of growth outcomes — formal models (with Trammell) show everything from modest acceleration to explosive growth, depending on whether AI substitutes for or complements labor. Held — now published (Annual Review of Economics 18, 2026), and the Anthropic model puts numbers on it: 2030 GDP +1.6%, +8.3% or +32.4% above baseline across the three branches.
• Empirical: Wages could fall even as output explodes — if AI is a near-perfect labor substitute, the economy grows but workers don't benefit. Drifted — mechanism held, valence revised by his own model: average wages of employed workers rise in all three branches through 2030 (+0.7% modest, about +10% extreme); losses concentrate on knowledge workers (−11.5% extreme) while non-knowledge wages jump (+34%), and total labor income in the extreme branch lands roughly where it would have without AI because employment falls (Darden Ideas to Action, 2026-09-15; Tabarrok, 2026-09-14). Outright collapse survives only as the full-automation endpoint of Korinek–Suh, which the 2030 model does not reach. The 2024 claim was aggregate; the 2026 claim is distributional.
• Empirical: Tax base collapses without reform — fiscal systems tax labor income; automation erodes that base, requiring new instruments (capital, robots, data). Held — elaborated into two stages: labor taxes give way to consumption taxes, then to taxing the AGI systems directly ("an optimal harvesting problem"); robot, compute and token taxes, sovereign wealth funds and windfall clauses each evaluated (Korinek & Lockwood, "Public Finance in the Age of AI: A Primer," NBER w34873, 2026).
• Normative: Policy must prepare before transformation hits — the speed and breadth may not allow gradual adjustment. Held and escalated: he co-organized "We Must Act Now: A Statement on AI's Transformation of the Economy" (2026-07-13) — 16 Nobel laureates, 200+ initial signers, now over 2,000 — and co-ran the IMF's scenario-planning workshop (2025-12-10/11), published as an IMF Note (2026-04-03).
• Empirical (new, 2025-11): Frontier-lab timelines are "not too wrong" — his "personal best guess" is years, not decades, though "nothing is inevitable" (Richmond Fed, Q4 2025). Live — ungraded; check under Predictions. Note he said this roughly five months before joining the lab whose timelines he was endorsing.
• Empirical (new, 2026-02): Firms may delay entry-level hiring on the expectation of rapid capability gains, depressing opportunity before any measurable job destruction (Ideas in Development, 2026-02-17). Held — six months later the Stanford Canaries update found employment of 22–25-year-olds in AI-exposed occupations 19% below their less-exposed peers, "primarily through reduced hiring of young workers rather than increased separations," with no economy-wide displacement (Stanford DEL, 2026-08).
• Normative (2025-04): Steer technical progress toward complementing workers rather than replacing them (Korinek & Stiglitz, INET WP 232; TIME100 citation). Unfalsifiable as a preference — noted, not graded; see Weak spots for the Says/Does tension.

Notable predictions — with falsifiable checks

• (2024–2025) If AI substitutes broadly for labor, wages and labor share will fall even as GDP grows. Drifted — labor-share half held (his 2026 model: −0.6 to −14.8 points by 2030); wage half revised to "knowledge-worker wages fall, average wages rise" (Darden, 2026-09-15). Precondition unmet: real GDP grew 1.5% annualized in Q2 2026 (BEA, 2026-08) and real average hourly earnings fell 0.1% year-over-year (BLS, 2026-09-11) — output is not exploding, so the conditional is untested. Check: by end-2030, if real GDP growth averages above 5%/yr, BLS nonfarm labor share falls below 49% AND the BLS real wage index for management/professional occupations declines; if growth stays under 3%, untriggered, not wrong.
• (2025) Current fiscal systems are unprepared for labor-share collapse. Held — the 2026 primer still describes advanced-economy tax systems as labor-income-dependent (NBER w34873); no G7 enactment of a compute, token or robot tax or an AI-funded sovereign wealth fund was found in this review (none found, not "none exists"). Check: any such enactment by end-2028.
• (Ongoing) The range of outcomes is much wider than policymakers appreciate. Unfalsifiable as written — rewritten: "no G7 finance ministry or the IMF runs a formal transformative-AI scenario exercise by end-2027." Already met the other way — the IMF ran one in December 2025 and published it in April 2026, co-organized by him. The premise is being falsified by his own success; retire the framing if a second G7 body follows.
• (2025-11) Lab timelines are "not too wrong." Live. Check: by end-2028, an AI system completes the majority of remote cognitive tasks at human quality on an independent benchmark (e.g., METR's 50%-success time horizon exceeding one working month); absent that, Drifted; absent by 2030, Wrong.
• (2026-09) Three-branch 2030 forecast (Anthropic Institute, v1.0). Per-branch checks, measured as change from the Q2-2026 BLS nonfarm labor share of 52.8% because his ~60% baseline is a different series: modest = labor share ≈52% and management/professional unemployment under 4%; substantial = ≈49% and 5–10%; extreme = ≈38% and above 15% (his extreme branch: overall unemployment 11.9%, knowledge-worker 17.9%). Whichever lands, the other two are graded Wrong for that horizon — the tooze rule, so the taxonomy can lose.
• (2026-01) Automating AI research: "the threshold for a growth take-off is within reach" (Davidson, Halperin, Houlden & Korinek, NBER w35155). Unfalsifiable as stated — rewrite or retire: proposed rewrite — sustained real GDP growth above 8%/yr in a G7 economy for two consecutive years by 2035.

Revealed behavior (Does)

• Took the Anthropic role (between 2026-02 and 2026-07) — time and paycheck now come from a frontier lab, and the first product of the move is a public scenario model plus interactive explorer: he used the lab to scale his scenario-planning program, not to build anything commercial.
• Spent 2025–26 building institutions, not writing op-eds: EconTAI (2025-08/09), the NBER Economics of Transformative AI workshop (2025-09), the IMF workshop (2025-12), the PIIE seat (2026-02), the 2,000-signature statement (2026-07).
• Four substantive papers in roughly twelve months (Annual Review, public-finance primer, explosive growth, Anthropic scenarios) — the cadence of someone who thinks the clock is short.
• Co-organized the statement with Brynjolfsson, Ajay Agrawal and Tom Cunningham (METR) — the persona this file once listed as his main "split" is now his co-signatory.

Feels

Urgency in a formalist's suit: he wants to be the economist who had the models ready before the shock, and fears a replay of the China-shock policy failure at ten times the speed. Wants to be true: that scenario planning gets institutionalized in time, and that "steer toward complements" is still a choice.

Hears

Frontier-lab timelines, now from inside Anthropic (Economic Index team, Chad Jones, Peter McCrory); the explosive-growth circle (Trammell, Davidson, Halperin — Open Philanthropy/Forethought lineage); Stiglitz on the direction of technical change; Brynjolfsson and Agrawal on measurement; the finance-ministry channel (G7 panel, IMF Fiscal Affairs). Almost no labor-movement or sociological input.

Sees

Two rooms almost no one else is in at once: Anthropic's internal usage data and capability roadmap before publication, and the G7/IMF side where fiscal responses get drafted. If a lab's private timeline and a treasury's scenario exercise ever need reconciling, he holds both documents.

Incentive map

Re-derived from scratch — v1 treated him as an unconflicted academic, and that status has inverted. Who pays him: Anthropic (employee, Head of Transformative AI Economic Studies, on leave from UVA); the council seat predates the hire by roughly a year and, per the Bengio and Wissner-Gross rules, should have been read as a leading indicator. Secondary: UVA (leave terms unpublished), PIIE and Brookings affiliated seats, NBER/CEPR. EconTAI's only named funder is UVA's College of Arts & Sciences; no external grant found. What he's selling: urgency (the statement), scenario planning as a service to ministries (IMF, G7), and the "steer toward complements" agenda. What he can't say now: "Anthropic's product is the substitution engine my models warn about," "lab timelines are marketing," or "the extreme scenario is implausible" — each cuts against his employer's valuation or his own program. Note the convenient alignment: the extreme branch is simultaneously a warning to policymakers and an advertisement for the lab's capabilities. That doesn't make the model wrong — Cowen calls the authors "sober, reasoned, scientific" (2026-09-10) — but thesis and payroll now point the same way. Rule-25 discount: high on timelines and magnitude; moderate on formal results (peer-reviewed, mostly written before the hire); low on the tax-base arithmetic, which is accounting, not forecasting.

Theories aligned with

• Post-labor economics (formal variant) — the two-thirds-of-income arithmetic and the public-finance primer are its rigorous spine
• Automation displacement — Korinek–Suh's race between automation and capital accumulation is the model version
• Entry-level work crisis — his 2026-02 hiring-delay mechanism, confirmed by Canaries 2026-08
• End of white-collar labor — the extreme branch only (knowledge-worker unemployment 17.9% by 2030)
• Exponential curves — the automated-research paper, threshold unfalsifiable as stated
• Universal basics — "we are going to need a new system of income distribution" is his standing line
• Growth theory under technological transformation; adjacent to Acemoglu/Autor labor economics but with explicit AGI scenarios

What he's reacting against

• Economic complacency about AI as "just another technology" — argues transformative AI breaks the assumptions of standard growth models (the Kaldor facts: long-run regularities like a stable labor share)
• Vague hand-waving about "AI will change everything" — insists on formal models with specific scenarios and testable implications
• Policy unpreparedness for a world where labor share collapses — most tax and welfare systems assume labor income remains dominant
• Both AI hype and AI dismissal — his models show the range is genuinely wide, making both extremes wrong
• NEW (2026): economists who wait for data before modeling — scenario planning, in his framing, is acting under uncertainty; the IMF exercise and the statement are its institutional form
• NEW (2026-09): reading "labor share falls" as "workers get poorer" — his own model separates the two

Where he overlaps / splits (with Rich)

• Overlaps with Acemoglu on labor-market concerns; splits on formalism — Korinek builds explicit transformative-AI growth models, Acemoglu focuses on institutional and task-level analysis. Held — now quantified: Korinek's extreme branch (17.9% knowledge-worker unemployment by 2030) is a world Acemoglu's modest-effects estimates do not contain.
• Overlaps with Susskind (A World Without Work) on labor displacement; splits on register — Korinek is formal economics, Susskind is policy narrative. Unchanged; Susskind is not on the bench.
• Overlaps with Cowen on AI as a major economic event; splits on timeline and magnitude — Korinek models transformative scenarios Cowen considers unlikely near-term. Held — Cowen on the 2026 model: "I will opt for something a modest bit more than modest" (2026-09-10). Independence note: both sit on the Anthropic council, so Cowen's praise is not a fully outside read.
• Splits with Brynjolfsson on complementarity — Brynjolfsson emphasizes human-AI collaboration, Korinek models the substitution endpoint. Drifted — they co-organized "We Must Act Now" (2026-07-13), and Korinek's hiring-delay mechanism is now the leading explanation for Brynjolfsson's own Canaries data. Analytically distinct, institutionally merged; per the Moonshots rule their mutual confirmation is discounted from here on.
• Overlaps with Frey on displacement-first — Canaries' 19% entry-level gap credits both; Frey's is the historical argument, Korinek's the expectations mechanism
• Overlaps with Amodei institutionally and on urgency; splits on the number — Amodei's "half of entry-level white-collar jobs" is Korinek's extreme branch, while Korinek's modest branch leaves knowledge-worker unemployment unchanged at 2.9%
• Overlaps with Aschenbrenner, Kokotajlo on short timelines and explosive growth; splits on method — they write scenarios as narrative, he writes them as calibrated models with a public explorer
• Overlaps with Mazzucato on steering technology's direction (via Stiglitz); splits on the lever — she wants mission-driven state investment, he wants tax design and scenario planning
• Overlaps with Shapiro on post-labor economics as destination; splits on rigor and on optimism about the transition
• Splits with Evans, Zitron on whether any of this is measurable — the Canaries confirmation of his hiring-delay mechanism is the bench's cleanest evidence against the "can't be measured" position

Track record

• Named TIME100 AI (2025-08) — recognition of field influence; Vox Future Perfect 50 (2024)
• Founded the Economics of Transformative AI Initiative at UVA (2025-08/09) — institution-building signals long-term commitment; ran the NBER Economics of Transformative AI workshop (2025-09)
• G7 High-Level Panel of Experts on AI (report to G7 finance ministers, 2024-12-19); testified to the US Senate AI Insight Forum on Workforce (2023-11-01) with the three-scenario framing that still structures his work — policy influence is real
• IMF workshop co-organizer (2025-12-10/11) — bridging academia and international policy institutions; output became an IMF Note (2026-04)
• Formal models are rigorous but rely on assumptions about AI capability that remain uncertain — the 2026 model's own caveat: "a stark simplification of a complex reality"
• NEW: PIIE nonresident senior fellow (2026-02-02); Head of Transformative AI Economic Studies, Anthropic (by 2026-07); co-organizer of the 16-Nobel statement (2026-07-13); lead author of the Anthropic Institute scenarios paper and explorer (2026-09)
• Scored this cycle: one mechanism claim confirmed at market level within six months (hiring delay → Canaries); one aggregate claim revised by his own model (wage collapse → distributional split); no outright empirical miss, because the transformative precondition has not arrived. Pattern: strongest on framing and institution-building; the quantitative forecasts are all still inside their windows

Empirical vs normative

• Empirical: growth models with specific parameter ranges; labor-share projections (−0.6 to −14.8 points by 2030); tax-base analysis; the hiring-delay mechanism; "lab timelines are not too wrong"
• Normative: policymakers should prepare for transformative scenarios; fiscal reform is urgent; "wait and see" is inadequate; steer progress toward complements; "take care of the losers in this transition"
• Unusually formal for the AI-society space — the models have testable implications under specified conditions, and as of 2026-09 carry published numbers and a public explorer anyone can re-run
• Conflict note (CONVENTIONS rule 25): Anthropic employee plus council seat — high discount on timeline and magnitude claims, moderate on formal results, low on tax-base arithmetic. The v1 file carried no discount at all.

Weak spots / open questions

• Models depend on assumptions about the AI capability trajectory — if AI complements rather than substitutes for labor, the dramatic scenarios don't materialize. Still true; the 2026 model makes it explicit by letting the user set the dials.
• "Transformative AI" is doing a lot of work as a concept — the models apply to a scenario that may or may not arrive. Still true; the market data (1.5% GDP growth, flat real wages) says it has not.
• Less engaged with cultural/philosophical dimensions — the frame is pure economics. Unchanged.
• Policy prescriptions (fiscal reform) are sound in principle but politically difficult — the gap between analysis and implementable policy is wide. Unchanged; the primer catalogs instruments without a political path for any.
• Early-career in this specific area (relative to Acemoglu/Autor) — influence is growing but institutional weight is still building. Wrong — graded against the file's own assessment: within 15 months he added TIME100 AI, a G7 panel report, an IMF scenario exercise, a PIIE seat, a frontier-lab department and a 2,000-signature statement with 16 Nobel laureates. He is the field's convening center, not its junior. The miss: we measured standing by seniority relative to Acemoglu/Autor instead of by who was routing money and signatures through him.
• NEW: the Says/Does tension — the "steer toward complements" economist now works for a lab whose product substitutes for cognitive work; no public reconciliation found. Watch whether his modest branch survives inside a company whose CEO markets the extreme one.
• NEW: the model can produce any wage sign — Trammell–Korinek shows wages rise or fall depending on returns to scale, resource bottlenecks and the direction of technical change. A framework that fits every outcome forecasts none; the per-branch checks above are the file's attempt to make it lose.
• NEW: the 2026 model excludes policy responses, business cycles, demand and financial disruptions, and catastrophic risk — by the authors' own statement. The extreme branch's 11.9% unemployment is a floor on disruption, not an estimate of it.
• NEW: write-ups of the 2026 numbers disagree (Darden quotes "+15%" for the extreme branch; Anthropic's page says GDP 32.4% above baseline by 2030 — likely annual rate versus cumulative). The file trusts Anthropic's page for levels and Darden's for the unemployment/wage split; read the technical PDF directly next review.

Rich's take

• (your synthesis here)

Delta log

2026-09-21 — v1→v2 migration + validation (bulk run)

• Grades: 7 Held / 3 Drifted / 1 Wrong (plus 2 unfalsifiable items rewritten, not graded, and 2 live predictions with checks attached).
• Wrong: the file's own weak spot "institutional weight still building" — falsified by PIIE (2026-02-02), the Anthropic department (by 2026-07-15), the 16-Nobel statement (2026-07-13) and the IMF Note (2026-04-03); sources linked in Track record.
• Drifted: "wages fall even as output explodes" and its prediction twin — his own 2026 model has average wages rising and losses concentrated on knowledge workers (Darden, 2026-09-15; Anthropic Institute, 2026-09; linked in Key claims); "splits with Brynjolfsson" — co-organizers of the same statement (econtai.org, 2026-07-13).
• Held: labor-share collapse as a conditional (Richmond Fed Q4 2025; BLS 2026-09-15 for direction); wide range of growth outcomes (Annual Review 2026); tax-base erosion (NBER w34873); prepare-before-it-hits (statement + IMF Note); fiscal systems unprepared (no enactment found); hiring-delay mechanism (Canaries, 2026-08); the Cowen split (MR, 2026-09-10).
• Most surprising delta: the bench's "unconflicted formal economist" became a frontier-lab department head inside the review window, and his first lab output is the most falsifiable thing he has published — a three-branch 2030 forecast with numbers and a public explorer. Conflict and testability rose together.
• Employer/role check (mandatory): job change — on leave from UVA as Head of Transformative AI Economic Studies, Anthropic (Darden News, 2026-07-15); PIIE nonresident senior fellow (2026-02-02). The 2026-06-16 review could not have caught a July item, but the February PIIE seat and the two January papers were public and absent — consistent with the May/June-2026-pass-is-defective rule. Incentive discount raised from none to high on forward claims.
• Dated post-review sources: Darden 2026-07-15, econtai.org 2026-07-13, Anthropic Institute 2026-09, MR 2026-09-10 and 09-14, Darden Ideas 2026-09-15, BLS 2026-09-15, Stanford DEL 2026-08.
• Tier: quarterly — a new employer, a dated 2030 forecast, quarterly BLS labor-share prints and a forthcoming NBER volume all move within a quarter; wake triggers cover the leave ending, model revisions, labor-share thresholds, tax enactments and Canaries reversals.
• Confidence unchanged at medium — the record is rich but the Anthropic start date and terms are unpublished, and every quantitative forecast is still inside its window.

Sources

• 2026 primary: Anthropic Institute — Economic Scenarios for Transformative AI, v1.0 (2026-09) · technical PDF (Korinek, Jones, Sacher, Cotter, McCrory) · "We Must Act Now" (2026-07-13) · Darden News — statement and Anthropic leave (2026-07-15) · PIIE release (2026-02-02) · IMF Note 2026/002 (2026-04-03) · IMF–EconTAI workshop (2025-12-10/11)
• Papers: Trammell & Korinek, Annual Review of Economics 18 (2026) · Korinek & Suh, "Scenarios for the Transition to AGI," NBER w32255 (2024-03) · Korinek & Lockwood, public-finance primer, NBER w34873 (2026) · Davidson, Halperin, Houlden & Korinek, NBER w35155 (2026) · Korinek & Juelfs, "Preparing for the (Non-Existent?) Future of Work," Oxford Handbook of AI Governance (2024) · Korinek & Stiglitz, "Steering Technological Progress," INET WP 232 (2025-04) · "Economic Policy Challenges for the Age of AI," MIT Press (2025) — all at korinek.com/research
• Interviews / talks: Richmond Fed Econ Focus (Q4 2025) · San Francisco Fed, "The Economics of Transformative AI" (2025-04) · Ideas in Development (2026-02-17) · Darden Ideas to Action (2026-09-15)
• Reactions: Cowen, Marginal Revolution (2026-09-10) · Tabarrok (2026-09-14)
• Market-level data used for grading: BLS, labor share 52.8% in Q2 2026 (2026-09-15) · BLS Real Earnings, August 2026 and BEA GDP Q2 2026 (linked in Predictions) · Stanford DEL, Canaries in the Coal Mine — August 2026 update (PDF)
• Roles / recognition: Anthropic Economic Advisory Council (2025-04-28) · TIME100 AI 2025 · G7 High-Level Panel report (2024-12-19) · Senate AI Insight Forum statement (2023-11-01, linked in Track record) · NBER Economics of Transformative AI workshop (Fall 2025)
• Bio pages (lagging — no leave shown as of 2026-09-21): korinek.com · UVA Darden directory · Wikipedia

migrated v1→v2 2026-09-21 · backup: _archives/anton-korinek.html.bak-20260921 · converts-from: people/anton-korinek.md · AI & Society domain