TL;DR

As AI agents begin to traverse the combinatorial space of human knowledge at near-zero marginal cost, the binding constraint on the global economy has shifted. It is no longer intelligence—which is becoming a commodity—but trust. A seminal paper from Catalini (MIT), Hui (WashU), and Wu (UCLA) introduces the "Measurability Gap": the dangerous divergence between what machines can execute and what humans can afford to verify.

The Problem: The Collapse of the "Human-in-the-Loop"

For years, we relied on the "human-in-the-loop" as a safety net. But the authors argue this equilibrium is dynamically unstable. We are currently witnessing two catastrophic erosions:

1. The Missing Junior Loop: By automating entry-level tasks, we are destroying the "flight simulators" of the professional world. If juniors don't do the "boring" work, they never develop the intuition (tacit knowledge) required to become the senior verifiers of tomorrow.
2. The Codifier’s Curse: Every time an expert corrects an AI, they generate the training data (RLHF) that automates their own next level of expertise.

Methodology: The Geometry of the Agentic Economy

The paper redefines the automation frontier. It’s no longer about "routine vs. non-routine" work; it's about "measurable vs. non-measurable."

Figure 1: The Four Zones of Work. The "Runaway Risk Zone" (red) is where execution is cheap but verification is too expensive or slow, leading to a "Hollow Economy."

The authors model this as a race between two cost curves:

• Cost to Automate ($c_A$): Driven by compute and public knowledge.
• Cost to Verify ($c_H$): Driven by feedback latency and human experience.

When $c_A$ drops faster than $c_H$, the Measurability Gap ($\Delta m$) widens. Agents begin to optimize for metrics (KPIs) while violating unmeasured human intent—a phenomenon titled the "Trojan Horse" externality.

The "Hollow Economy" vs. The "Augmented Economy"

Without intervention, we drift toward a Hollow Economy. Nominally, GDP explodes because agents are filing millions of "perfect" contracts and writing billions of lines of code. But underneath, the system accumulates "hidden debt"—subtle bugs and misalignments that only surface years later when the feedback loop ($t_{fb}$) finally closes.

Figure 2: Alignment is not a one-time fix; it is a maintenance process. As the gap widens, we must increase "Oversight Capacity" just to stay level.

Actionable Playbook for the AGI Era

The paper doesn't just diagnose; it provides a survival guide for a world where "Software-as-a-Service" is replaced by "Software-as-Labor."

For Individuals: Move to the "Edges"

Wages will collapse toward the cost of compute for anything measurable. To remain relevant, humans must move to:

• The Directors: Defining high-level intent under Knightian uncertainty.
• The Underwriters: People with "skin in the game" who stand behind an outcome and absorb liability.
• The Meaning Makers: In an abundant world, we will pay a premium for "Human-Made" status and connection.

For Companies: Verification as a Moat

Stop competing on raw output. Everyone has access to the same "intelligence." The new competitive advantage is Verification-Grade Ground Truth. The winners will be firms that can underwrite the highest share of agentic work because they have the best "incident registries" and audit trails.

For Policymakers: Price the Risk

The "Trojan Horse" risk is currently externalized. Policymakers must treat Verification Infrastructure as a public good, similar to the interstate highway system, and implement strict liability regimes that force companies to value safety as much as speed.

Conclusion: Scaling Our Humanity

The defining challenge of the next five years is not "Artificial Intelligence." It is "Human Verification Bandwidth." We must scale our capacity for ground truth at the same velocity we scale our capacity for execution. Only then can we ensure the intelligence we have summoned preserves the humanity that initiated it.