What becomes scarce when AI execution grows cheap?

The paper argues that once autonomous agents can act with "broad agency rather than narrow instructions," the binding constraint on economic growth is no longer intelligence but "human verification bandwidth: the scarce capacity to validate outcomes, audit behavior, and underwrite meaning and responsibility when execution is abundant." It models the transition toward AGI as "the collision of two racing cost curves": an exponentially decaying Cost to Automate (cA), driven by compute and accumulated knowledge, against a biologically bottlenecked Cost to Verify (cH), bounded by human time and embodied experience. The gap between the two, the Measurability Gap (Δm), determines what share of agentic output is actually verified and thus productive — the "verifiable share of deployment" (sv) — versus merely "plausible."

The model extends Romer's (1990) endogenous growth theory but reorients its locus of innovation around task automation rather than capital-variety expansion, decomposing Effective Labor into Measurable Capacity (Lm, where agentic labor substitutes for human execution) and Non-Measurable Capacity (Lnm, the human steering and verification that directs it). Unverified agentic output does not simply fail to help; it "leaks into the economy as a 'Trojan Horse' externality (XA) that actively parasitizes the general capital stock and crowds out human consumption." Because tasks, not skills, now sort what can be automated — "any task that can be reduced to a metric can be industrialized (Catalini et al., 2025)... regardless of the prestige or complexity" — the paper names this dynamic "measurability-biased technical change," replacing the older skill-biased framing. It also argues the human capital needed to verify is self-destroying: automating entry-level measurable work removes "the training ground required to build future experts," a failure it calls "the Missing Junior Loop."

This supplies a formal mechanism for the institutional claim in What makes accountable judgment scarce when AI cognition is cheap?, which names "accountable judgment" as the scarce asset without modeling why it stays scarce; here the scarcity follows from a biologically inelastic cost-to-verify curve racing against an exponentially falling cost-to-automate curve. It also gives a labor-economics channel to Does incremental AI replacement erode human influence over society?: where that argument treats loss of human labor participation as eroding implicit societal alignment, this paper explains one way labor gets lost — entry-level measurable work is automated first, destroying the apprenticeship pipeline that produces future verifiers. It extends the task-based automation lineage behind Does automation raise or lower the skills that remaining work demands? by replacing the routine-versus-non-routine axis with a measurable-versus-non-measurable one, and it reads as a more pessimistic counterpart to Are AI feedback loops strong enough to sustain recursive self-improvement?, which finds today's capability feedback loops still weak, where this paper treats the automation-cost curve as already racing ahead of verification capacity.

The excerpt is an extended abstract plus method and discussion sections of a theoretical paper; it states a formal model and strategic implications, not an empirical measurement of any real firm, occupation, or deployment. The cost curves (cA, cH), the Measurability Gap, and the verifiable share (sv) are modeled constructs in this excerpt, not quantities estimated from data. What follows at the strength the excerpt supports is a conceptual warning — that pressure to automate creates an incentive for "unverified deployment" whose hidden risk accumulates — rather than a forecast of when or how severely any specific economy or industry hits this constraint.