AI and the Complementarity Theory of Economic Growth
How should we model AI’s effect on economic growth? The standard Solow growth framework — one aggregate production function, one kind of capital, exogenous technology — was built for a world where the “intelligence sector” never takes sudden swings. AI breaks that assumption, and economists have responded with three very different frameworks. They disagree not mainly about AI’s capabilities but about what binds: whether the bottleneck to growth is intelligence itself, the complementary inputs intelligence needs to be useful, or the tasks where intelligence can actually be applied. This page maps the three positions; the disagreement is live and unresolved.
Cowen’s two-factor model: Intelligence × Polanyi knowledge
Tyler Cowen’s September 2026 sketch replaces Solow’s single capital stock with two factors of production:^[Tyler Cowen 2026 — A simple model of AI-aided economic growth
- Intelligence — formal smarts: playing chess, proving theorems, scoring well on evals. Humans and AIs both supply it, and AI has delivered a huge positive shock to the total stock.
- Polanyi knowledge — Michael Polanyi’s tacit, inarticulable knowledge: of time and place, custom and habit, how a particular office actually works. The knowledge Hayek argued could never be centralized. Humans specialize in it; AI can assist but cannot yet be “brought into your office to figure out how that office works” in the human rather than mechanistic sense.
Output is produced by combining the two, and they are near-complements — not quite Leontief (fixed-proportion), but close. A dose of AI-drenched technocratic knowledge does not solve an office’s norms problem, and can make it worse by empowering rent-seekers.^[Tyler Cowen 2026 — A simple model of AI-aided economic growth
The model’s predictions follow from complementarity alone:
- A shock to Intelligence raises marginal returns, employment, and real wages in the Polanyi sector, because those inputs are now relatively scarce. The gains are slow (tacit knowledge can’t be boosted quickly — “messy by its nature,” citing Luis Garicano) but durable, since absorbing existing AI advances takes a long time.
- There is transitional unemployment in the Intelligence sector once “centaur” models (human+AI teams) fade — as happened in chess, where human grandmasters were briefly essential partners and then not. Mathematicians currently doing the prompting are a live centaur case.
- Commandeering the Intelligence sector buys less power than it appears — without the complements, control of the AI stock changes little at first.
- Growth is robust but not explosive: shocking AI advances, a fine job market, and growth that rises only as the Polanyi sector slowly catches up. Cowen notes this matches 2026 data.^[Tyler Cowen 2026 — A simple model of AI-aided economic growth
The model also explains why Solow worked for so long: the Intelligence sector rarely moves fast, so the ratio between the two factors stays roughly constant in the short run — until it doesn’t.
The growth-theory spectrum: where three frameworks disagree
Cowen’s sketch sits between two more formal treatments that bracket the range of credible outcomes.
Acemoglu: task-based, modest effects
Daron Acemoglu’s “The Simple Macroeconomics of AI” (NBER WP 32487, April 2024; published in Economic Policy 2025) works through a task-based model where AI’s microeconomic effects are cost savings at the task level.^[Daron Acemoglu 2024 — The Simple Macroeconomics of AI By a version of Hulten’s theorem, aggregate gains are bounded by (fraction of tasks exposed) × (average task-level cost savings). Using existing exposure estimates (about 20% of US labor tasks exposed to AI; about 23% of those profitably automatable) and cost-savings estimates from early studies, Acemoglu computes total factor productivity gains of no more than ~0.7% over ten years — and argues even that is an overestimate, because the early evidence comes from easy-to-learn tasks while future gains must come from hard-to-learn, context-dependent tasks where there are no objective outcome measures to train on. His updated bound: TFP +0.53%, GDP +0.90% over ten years — an order of magnitude below the transformative scenarios.^[Daron Acemoglu 2024 — The Simple Macroeconomics of AI
Acemoglu’s framework is essentially Cowen’s skepticism formalized: the “hard-to-learn tasks” that cap his estimates are exactly the tasks saturated with tacit, context-dependent knowledge that Cowen assigns to the Polanyi sector. He also warns that some new AI-generated tasks (engagement-maximizing algorithms, manipulation) may have negative social value, so GDP can rise while welfare falls.
Trammell & Korinek: automation of production and R&D
Philip Trammell and Anton Korinek’s “Economic Growth under Transformative AI” (GPI Working Paper 8-2020; NBER WP 31815; August 2025 revision) surveys the full possibility space rather than committing to a point estimate.^[Philip Trammell and Anton Korinek 2025 — Economic Growth under Transformative AI Their key distinction:
- Automating production (machines can self-replicate — capital substitutes for labor): robustly raises growth and drives the labor share toward zero, breaking both Kaldor facts (constant per-capita growth, constant labor share) that have characterized industrial economies for two centuries.
- Automating R&D (machines can self-improve): accelerates the transformation but may not produce it in isolation — it depends on whether automated research is highly parallelizable and whether it needs inputs that scale with economic activity rather than with compute.
- Wages = exploding output × plummeting labor share: may rise or fall depending on returns to scale, natural-resource constraints, and the direction of technical change. Genuinely uncertain even in sign.^[Philip Trammell and Anton Korinek 2025 — Economic Growth under Transformative AI
The disagreement in one sentence
Acemoglu says AI is bounded by the set of tasks where it delivers measurable cost savings (small); Cowen says AI is bounded by the supply of tacit complementary knowledge (large and durable but slow); Trammell & Korinek say that if AI crosses the threshold of full production automation, both bounds break and growth accelerates dramatically — with the labor share, not necessarily wages, as the casualty. The frameworks differ less on what AI can do today than on whether the complementarity constraint is permanent (Cowen/Acemoglu) or itself automatable (Trammell & Korinek).
The capital-supply margin
Fernández-Villaverde’s “Saving for the Machines” (PIER WP 26-018, October 2026) adds a dimension that cuts across the three frameworks above: holding the automation technology path fixed, capital supply governs how large the boom gets. The same technology path delivers 11.7% annual growth when households choose their saving rate optimally and 2.2% when the saving rate stays at its calibrated 2025 level of 23% — automation raises the capital share (one-third toward one under full automation), so the saving rate rises endogenously or the boom shrinks. The model sits in the Acemoglu–Restrepo task framework, and its fiscal aftermath (real yields jump on impact; adopter and bystander debt arithmetic) is worked through in saving-for-the-machines.1
The 2026 empirical record: reallocation on the hiring margin
Ahn & Carollo (Federal Reserve Board, September 2026) supply the first systematic flows-based measurement of the LLM-era US labor market, matching CPS microdata to occupation-level AI exposure scores (Eloundou et al. 2023) and an industry-level AI-adoption proxy from Lightcast job postings, with JOLTS vacancies allocated to occupations via posting composition.2
Their headline results read like complementarity-constrained adjustment operating in the wild:3
- In cross-section, high-exposure/high-adoption workers have lower unemployment than low-exposure/low-adoption workers — driven by lower separation rates (the familiar skill gradient, not an AI effect).
- The post-LLM changes concentrate on the hiring margin: the high-exposure/high-adoption group’s unemployment rose ~1 percentage point from end-2022 to July 2026 (vs. ~0.3pp for the low/low group), driven primarily by declining job-finding rates — not rising separations. Job-to-job switching also fell for the exposed groups.
- Within-job activity switching rose noticeably for the same workers: employers are reorganizing tasks inside continuing jobs rather than firing.
- The pre-LLM positive association of AI intensity with nominal wage growth and hours attenuated markedly after LLMs arrived.
- Recovering the natural rate of unemployment from trend inflow/outflow components by exposure group, they estimate it has risen ~0.1–0.2 percentage points since the introduction of LLMs — “albeit with considerable uncertainty.”
Read against the three frameworks above: no mass displacement of incumbents (Acemoglu’s modest-effects world holding so far), pressure concentrated where exposure meets actual adoption, and adjustment through within-firm task reorganization — the Polanyi-sector absorption working itself out inside companies. The displacement-through-hiring-freeze shape also rhymes with the bootmaking record in technological-unemployment-history: adjustment lands on the entry margin (job-finding), not on incumbents. The paper is descriptive, not causal — the AI-attributed signal is the differential between exposure groups, not the common post-2022 hiring slowdown — and the natural-rate estimate carries the authors’ own considerable-uncertainty flag.4
Connection to the historical record
The task-and-complementarity framing gets empirical support from technological-unemployment-history: Vipond’s bootmaking study showed mechanization displaced jobs slowly, through entry collapse rather than incumbent displacement — consistent with a world where new technology must wait for its complements (new firms, new skills, new locations) to be built. The historical record also warns that the early decades of each industrial revolution are the most disruptive, and that adjustment costs fall on new entrants — the cohort whose “Polanyi knowledge” has not yet been formed.
Open questions
- Is tacit knowledge a permanent bottleneck, or just a slow one? If AI systems accumulate contextual knowledge through persistent deployment (memory, fine-tuning on firm-specific data), the Polanyi sector could itself be partially automated — shifting the economy toward the Trammell–Korinek regime.
- Does the hiring-margin decline for AI-exposed groups stabilize as within-firm task reorganization completes, or accelerate if the centaur fade arrives? Ahn & Carollo’s within-job activity-switching series is the leading indicator to watch.
- Do the centaur transition dynamics in chess (brief human indispensability, then irrelevance) generalize to mathematics and other cognitive professions? See ai-mathematical-practice for the live case.
- If the labor share collapses while output explodes, do existing redistribution institutions survive the political economy of the transition? See post-scarcity-economics and universal-basic-income.
Cross-domain connections
- knowledge-problem — Hayek’s 1945 statement of the constraint Cowen’s Polanyi factor inherits: knowledge of particular circumstances of time and place that cannot be centralized, and the price system as the channel that uses it anyway
- price-of-intelligence — the price series any AI growth model consumes; Zhu’s token-vs-task divergence means the deflator choice can flip measured real growth’s sign
- ai-as-normal-technology — the worldview this page’s evidence base currently supports: slow diffusion, hiring-margin adjustment, and modest measured impacts are what a “normal” general-purpose technology looks like in the data
- technological-unemployment-history — the empirical record of what labor displacement actually looked like; the bootmaking “entry collapse” mechanism is the historical version of complementarity-constrained adjustment
- ai-mathematical-practice — the community-level response to AI in one intelligence-sector profession; Tao’s “centaur” observations are exactly Cowen’s transitional-intelligence-employment case
- ai-research-productivity — the micro-level distributional evidence behind this page’s complementarity: inside economics research, AI narrows research-assistant gaps and widens seniority gaps, matching the execution/conceptual split
- scion-agent-scale-source-control — Anthropic’s account of engineer roles in agent-scale infrastructure work: Claude writes the code, engineers supply design tradeoffs and the veto (“the role of the engineers was to tell Claude ‘no’”), the execution/conceptual split in operational form
- post-scarcity-economics — the theoretical endpoint if the Trammell–Korinek transformative scenario fully materializes
- bullshit-jobs — a different critique of the link between employment and productive output
- effective-altruism — the movement whose philanthropic arm coined “transformative AI” (Karnofsky 2016) and funded the GPI environment where the Trammell–Korinek survey originated
Sources
- Tyler Cowen 2026 — A simple model of AI-aided economic growth
- Philip Trammell and Anton Korinek 2025 — Economic Growth under Transformative AI
- Daron Acemoglu 2024 — The Simple Macroeconomics of AI
- HieJoo Ahn and Nicholas A. Carollo 2026 — Artificial Intelligence and Labor Market Reallocation
- Jesús Fernández-Villaverde 2026 — Saving for the Machines