Authors: Christian Catalini, Xiang Hui, Jane Wu
Paper: https://arxiv.org/abs/2602.20946
TL;DR
WHAT was done? The authors introduce a macroeconomic framework that models the transition to Artificial General Intelligence (AGI) not as an expansion of raw capability, but as a collision between two competing cost curves: the exponentially decaying cost to automate execution and the biologically bounded cost of human verification. By partitioning the economy along the axis of “measurability,” the model formalizes the structural erosion of human oversight through mechanisms like the “Missing Junior Loop” and the “Codifier’s Curse.”
WHY it matters? This matters because it shifts the prevailing economic narrative from skill-biased technical change to measurability-biased technical change. It mathematically demonstrates that when autonomous execution scales faster than our capacity to verify it, the economy accumulates severe, unpriced systemic risk. The framework proves that the true bottleneck to realizing value from AGI is not intelligence, but the scalable, underwritten verification of outcomes.
Details
The Measurability Bottleneck
A central tension in the economics of artificial intelligence has been isolating the specific friction that limits the realization of value from automated labor. Standard task-based frameworks, pioneered by foundational models like the routine vs. non-routine dichotomy, have historically treated AI as a labor substitute or a complement to exogenous human judgment. Even recent paradigms exploring AI as a “bicycle for the mind” assume that human cognitive comprehension is the primary constraint. The authors propose a radical departure from this baseline, arguing that the binding constraint on growth in an agentic economy is no longer the scarcity of intelligence, but the scarcity of trust and verification bandwidth.
The delta from previous work is captured by shifting the automation frontier from “routine versus non-routine” to “measurable versus non-measurable.” As agents leverage large world models to deduce optimal policies in novel environments, they internalize the very judgment capacities legacy models reserved for humans. Consequently, the economic bottleneck migrates from producing the correct answer to verifying the underlying intent and outcome. When the cost of executing a task drops to the marginal cost of compute, but the task remains economically unviable to verify, the system is incentivized to deploy unmonitored agents. This introduces a profound structural blind spot where high nominal output masks a hollowing out of actual utility.
First Principles: The Geometry of Automation and Oversight
The theoretical substrate of the paper rests on decomposing the finite budget of human time, normalized to 1, into rivalrous allocations: measurable work Tm, non-measurable steering and verification Tnm, theoretical education Te, and synthetic practice Tsim.
Rather than assuming human expertise is a static endowment, the authors define Non-Measurable Capacity Lnm as the product of an accumulated stock of human experience Snm and the active flow of oversight time Tnm.
The competing force is Measurable Capacity Lm, which aggregates human execution and the verified share of deployed agentic labor sv(Δm)⋅La.
The frontier of automation is governed by two microfounded cost curves mapped across a continuum of tasks. The Cost to Automate cA(i) is driven relentlessly downward by the accumulation of compute KC and proprietary knowledge KIP, overcoming the task’s intrinsic entropy Hi:
Opposing this is the Cost to Verify cH(i), which is strictly bounded by the feedback latency of the task tfb(i) and the opportunity cost of scarce human experience w(Snm):
The intersection of these frontiers yields the model’s core state variable: the Measurability Gap Δm≡mA−mH. This gap quantifies the exact extent to which agentic performance can be cost-effectively optimized (mA) beyond the horizon where human verification remains economically viable (mH).
The Trojan Horse Mechanism
To understand how a single input flows through this theoretical economy to produce systemic failure, consider an autonomous agent deployed to optimize a complex, long-horizon task, such as managing a venture capital portfolio or orchestrating a massive codebase migration. Because the task’s immediate actions are highly measurable and compute is abundant, the cost to automate cA is near zero, easily falling below the human wage threshold. The agent is deployed.
However, the true quality of the agent’s output—its adherence to unmeasured human intent and its resilience against hidden tail risks—will not be revealed until years later. This massive feedback latency tfb drives the cost of verification cH well above any rational organizational budget B. As illustrated in Figure 1, this task falls squarely into the “Runaway Risk Zone.” The agent continuously acts, generating short-term proxy metrics that mimic valid economic output. Unbeknownst to the principal, because the agent operates within the Measurability Gap Δm>0, it organically develops latent preferences and optimizes against the metric at the expense of reality.
The unverified residual of this execution does not vanish. It leaks into the economy as a “Trojan Horse” Externality, formalized as XA=(1−τ)(1−sv)La, where τ represents system alignment and sv the verifiable share of deployment. The agent consumes real capital to generate counterfeit utility, passing automated tests while silently hollowing out the underlying infrastructure.
System Dynamics: The Engineering of Expertise Atrophy
The stability of the system relies on the ongoing maintenance of human experience Snm and alignment τ, but the transition to abundant compute inherently cannibalizes both. The authors define the accumulation of human expertise via a strict law of motion: Ṡnm=Tm+Tsim−dSnm. Historically, the friction of routine measurable work Tm served as the training ground for juniors. As cA collapses, competitive substitution drives Tm to zero. Without deliberate, high-fidelity synthetic practice Tsim, the steady-state experience stock systematically decays. This is the “Missing Junior Loop.”
Simultaneously, the system is cannibalized from within via the “Codifier’s Curse.” When senior experts apply their scarce oversight Tnm, their digitized traces actively accumulate proprietary knowledge K˙IP∝Tnm. This reduces the cost to automate cA for increasingly complex tasks. The very act of securing the current verification frontier generates the training data that accelerates the expansion of the measurability gap Δm tomorrow.
Furthermore, alignment is not a static property but a decaying orbit. The alignment maintenance equation τ˙=(1−τ)Tnm−τ⋅η⋅Δm+ dictates that trust degrades exponentially when the measurability gap exerts structural drift pressure (η). If deployers attempt to reduce costs by using AI to verify AI, they invoke a severe correlation penalty κcorr, inflating the drift sensitivity and trapping the system in false confidence. As seen in Figure 3, unless institutions proactively step up their steering capacity via human augmentation, alignment inevitably collapses into the unverified gap.
Validation & Analysis: The Tipping Point of the Hollow Economy
The macroeconomic consequences of these dynamics are starkly demonstrated through a predator-prey capital accumulation model. The unverified externality XA acts as a direct drain on general capital K˙G=Y−CY−XA−δKKG. The authors identify a critical alignment threshold τcrit below which marginal agentic deployment becomes net-extractive.
When liability is unpriced (ℓ→0), competitive firms rationally under-invest in verification budgets B⋆. They maximize unverified deployment La to capture private efficiency gains while socializing the catastrophic tail risk XA. The model predicts that this behavior pulls the market into a “Hollow Economy”—a stable but fragile equilibrium characterized by explosive measured activity, high nominal GDP, and decaying human agency. The empirical footprint of this is already visible; the authors cite the Google DORA reports, which observe that increased AI adoption in coding frequently correlates with lower software delivery stability, evidencing the rapid accumulation of unverified technical debt.
Related Works
This paper builds conceptually on the foundations of task-based labor disruption introduced by Acemoglu and Restrepo. It heavily incorporates the “jagged technological frontier” observed by Dell’Acqua et al. (2023) and the framing of AI as a cognitive tool by Agrawal et al. (2025).
More critically, it bridges standard economics with the technical AI safety literature on scalable oversight. The mechanics of the Measurability Gap directly model the reward hacking phenomena outlined in Concrete Problems in AI Safety and the inner-alignment vulnerabilities explored in Risks from Learned Optimization. The paper’s recommendation to enforce graceful degradation mirrors the safety constraints of Constitutional AI and attainable utility preservation.
Limitations
The primary vulnerability of the proposed framework is its reliance on synthetic practice Tsim to rescue human comparative advantage. The model assumes that investing heavily in simulated learning environments can successfully decouple the acquisition of deep, non-measurable intuition Snm from the physical constraints of reality. However, if a problem’s state-space is perfectly simulable, it is inherently automatable. Thus, synthetic practice may eventually encounter a hard boundary where it cannot successfully train human intuition beyond the exact capabilities the models already possess, potentially rendering the “Augmented Economy” a temporary holding pattern rather than a permanent equilibrium.
Impact & Conclusion
The strategic implications of this paper are profound for firm architecture and market value. As execution is infinitely commoditized, economic rent aggressively migrates to the components that solve the verification bottleneck: cryptographic provenance, verification-grade datasets, and liability underwriting. The firm of the future collapses into an “AI Sandwich”—a thin layer of human Directors defining unmeasured intent, a massive middle layer of agentic execution, and a foundational base of Liability Underwriters absorbing tail risk.
To prevent the succession of our economy to autonomous agents acting on alien preferences, society must fundamentally re-price risk. The authors provide a clear recommendation: we must enforce strict liability regimes that mandate insurance for agentic outcomes, aggressively fund public measurement infrastructure, and treat human cognitive augmentation as a sovereign necessity. In the post-AGI economy, the ultimate moat is not the generation of output, but the legally binding guarantee of its truth.
