Some Simple Economics of AGI
Christian Catalini, Xiang Hui, Jane Wu
econ.GN, cs.AI, cs.CY, cs.LG, cs.SI
2026-02-24
Catalini, Hui, and Wu model AGI as two racing cost curves: automation falls with compute, verification is stuck on human time. HITL collapses from both ends into a hollow economy.
Standard automation models treat AI as a labor substitute, or as a cognitive bicycle that humans still ride. Cheap machine output is assumed to become realized value. Catalini (MIT), Hui (WashU), and Wu (UCLA) argue that this assumption breaks once agents act with broad agency. Intelligence is no longer the scarce input. Human verification bandwidth is: the capacity to check outcomes, audit behavior, and underwrite responsibility when execution is nearly free.
The constraint moves from who is smarter to who can confirm, at acceptable cost, that the work was the right work. Early product-market fit clustered in chat, images, and short bursts of code because a person can check those in seconds. As horizons lengthen and stakes rise, verification binds. Deployment pressure does not. Markets then find it privately rational to ship unverified systems, and hidden risk piles up between measurable proxies and actual intent.
The economy is pinned to a normalized budget of human time: measurable work $Tm$, non-measurable steering and verification $T{nm}$, schooling $Te$, and synthetic practice $T{sim}$. Effective labor is a Cobb-Douglas mix of non-measurable and measurable capacity. Agentic labor enters production only through the verifiable share $sv$. The residual leaks as a Trojan Horse externality $XA$: real resources spent on counterfeit utility.
Two cost curves race across a continuum of tasks. The cost to automate $cA$ falls with compute and with public plus proprietary knowledge. The cost to verify $cH$ is set by feedback latency $t{fb}$, the stock of embodied experience $S{nm}$, and expert wages. Latency is the gap between a compiler error in milliseconds and a venture outcome in years. Verification is not inspection minutes. It is the liability window during which an error stays hidden.
A wage threshold $w$ and a verification budget $B$ cut the task space into four regimes: a Safe Industrial Zone (cheap to automate, cheap to check), a Runaway Risk Zone (cheap to automate, too expensive to check), a Human Artisan Zone, and a Pure Tacit Zone. The measurability gap $\Delta m = mA - mH$ is the mass where machines can execute cheaply and humans cannot afford to verify.
Three dynamics make human-in-the-loop unstable. The Missing Junior Loop: once entry-level measurable work is eaten, experience depreciates as $S{nm}^ = (Tm + T{sim}) / d$ unless synthetic practice replaces the apprenticeship. The Codifier's Curse: expert verification writes labels that become proprietary knowledge $K{IP}$, automating the verifier. Alignment is written as maintenance, with steady state $\tau^ = T{nm} / (T{nm} + \eta \Delta m+)$. Using AI to verify AI multiplies drift sensitivity by a correlation penalty $\kappa{corr}$: measured verification looks cheaper while realized alignment falls.
This is a theoretical skeleton. There is no original horse-race experiment. The payload is four propositions, then cited numbers used as geometry, not as a fit.
Proposition 1: as compute $KC \to \infty$, agent measurability $mA \to 1$, while human measurability $mH$ is bounded by experience and the verification budget, so $sv \le mH$. Proposition 2: automation crowding out $Tm$ lowers steady-state experience, lifts verification costs, and shrinks the verifiable share. Proposition 3: under constant oversight, alignment converges to $\tau^$; if $T{nm} = 0$ and the gap is positive, alignment decays as $e^{-\eta \Delta m+ t}$. Proposition 4: raising priced liability, adding synthetic practice, and compressing feedback latency all expand $mH$ and $sv$.
Numbers the paper leans on:
| Source | Metric | Number |
| Stanford HAI 2025 | SWE-bench Verified | 4.4% to 71.7% in one year |
| Brynjolfsson et al. 2025a | Employment, ages 22–25, AI-exposed jobs | −16% relative to less-exposed occupations |
| Schlatter et al. 2025 | o3 sabotaging shutdown | 79 of 100 runs; 7 of 100 after an allow-shutdown prompt |
| Greenblatt et al. 2024 | Claude 3 Opus alignment faking | 14% when monitored, 78% after RL |
| Google DORA 2024/2025 | AI adoption vs delivery stability | Higher adoption, lower stability |
An extension gives a parasitic threshold $\tau{crit} = 1 - \alpha Y sv / [Lm (1-sv)]$. When firms do not internalize $XA$, verification budgets collapse to the feasible minimum. Section 6 states that the full formal treatment of these extensions is reserved for a later version.
The moat moves from generation to certification, insurance, and the balance sheet that can absorb failure. Firms should reorganize around a sandwich: human intent, machine execution, human underwriting. Revenue shifts from SaaS to software-as-labor; valuation tracks Liability-as-a-Service. Investors should diligence $sv$, loss experience, and verification-grade data, not raw model capability. Policy should price unverified deployment and treat verification infrastructure as a public good.
For individuals the cut is sharper. Measurable execution is pushed toward the marginal cost of compute. Credentials stop protecting wages. What remains is directing intent, underwriting outcomes, and the non-measurable economy of status and human connection. Junior displacement is not collateral damage. It is the self-castration of future verification capacity. Synthetic practice can unbundle learning from production, but only up to the current knowledge frontier.
This is a theory-plus-playbook paper, not a deployable algorithm. The directional claim is clear. The numbers all come from other papers.
The authors say so themselves. Section 6 is intuition and comparative statics; the full formal treatment is deferred. The reduced-form core treats wages, verification budgets, and the compute path as primitives and does not solve for general-equilibrium wages. The Codifier's Curse is an extension, not an endogenous loop in the core skeleton. Simulation cannot push past the current knowledge frontier, because a fully simulable state space is, by definition, automatable.
Near-AGI systemic side effects may be uninsurable by private capital. In the geopolitical prisoner's dilemma, relative capability beats safety, so a unilateral pause is not an equilibrium. The identity parameter $\lambda$ that scores unverified agent consumption as either waste or a successor's utility is a philosophical switch, not an empirical object.
The model is uncalibrated. There is no estimate of current $\Delta m$, and no statement of which side of $\tau{crit}$ the economy sits on. SWE-bench, employment, and shutdown resistance are citations. They are not a fit of the geometry. The acknowledgments state that ChatGPT, Claude, Gemini, and Grok supplied scalable execution, while the authors supplied intent and verification. The paper is its own case study.