Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage -- by Ilse Lindenlaub, Ryungha Oh, Maria Alejandra Rodriguez, Laura Veldkamp

URL SCAN: Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage
FIRST LINE: We document and explain the gap between measures of AI exposure and measures of AI adoption in the workplace.

THE DISSECTION

This is a NBER working paper that performs a narrow empirical correction to a flawed conceptual framework. It finds that "AI exposure" — how much AI could replace a worker's tasks — predicts actual "AI adoption" only weakly (14% of cross-occupation variance). Their alternative: a comparative advantage framework that factors in worker-specific productivity, wages, and task-specific AI implementation costs. This gets them to 60%.

On its own terms: competent empirical work. As a contribution to understanding AI's labor-market trajectory: dangerously misleading in its framing.

THE CORE FALLACY

The paper treats AI adoption as a comparative advantage equilibrium problem — an optimization exercise where employers calculate whether AI beats a specific worker at a specific task, net of costs.

This is a static framework applied to a fundamentally dynamic discontinuity.

The DT lens exposes what the paper cannot see: comparative advantage for human labor is not a stable property. It is an asset under liquidation. As AI capability advances on a 12–18 month doubling curve, the "comparative advantage" calculation is not a steady-state optimization — it is a countdown. Every quarter the denominator shifts. Today's wage-attractive worker becomes tomorrow's unnecessary middle layer becomes the quarter after's archival curiosity.

The paper's finding that comparative advantage explains ~60% (not 100%) of adoption variance is not a nuanced empirical result. It is a measurement of institutional friction, implementation lag, and bureaucratic inertia — the very lag defenses the DT framework identifies. They are measuring the noise on the signal, not the signal.

HIDDEN ASSUMPTIONS

1. Worker productivity is the relevant frame. The paper treats human labor as the structural baseline and AI as a challenger. The DT framework inverts this: AI is the baseline, human labor is the anomaly being priced out.

2. Wage differentials drive the adoption decision. The framework implies that where humans are cheap relative to AI implementation costs, humans persist. This ignores that AI costs fall at a logarithmic rate while human wages (in aggregate) do not. The comparison is asymmetric by design.

3. Revealed preference from current adoption is a valid guide to future dynamics. The "user costs" inferred from current adoption patterns embed the very institutional inertia the paper is trying to explain. Extrapolating from a lag-weighted equilibrium tells you nothing about the inflection point.

4. Occupation boundaries are meaningful unit of analysis for AI. AI does not respect SOC codes. A framework built on cross-occupation variance is measuring the speed of managerial dysfunction, not the capability trajectory of the technology.

5. The 60% R² is a success metric. It is actually a confession: 40% of AI adoption is explained by factors other than economic rationality. That 40% is friction. Friction delays. It does not prevent.

SOCIAL FUNCTION

Institutional Delay Theater. This paper's function is to tell establishment economists and HR strategists that the AI transition will be slower, more nuanced, and more manageable than raw exposure metrics suggest. It is academic cover for the lag defense. "See? It's not as simple as exposure models claim. There's a rational optimization process. Institutions have time to adapt."

This is a lullaby dressed in regression tables.

THE VERDICT

The paper provides a useful empirical correction: yes, exposure ≠ adoption, and yes, comparative advantage matters in the transition. But the frame — that we are observing a rational optimization process that will result in gradual, task-level displacement — is the intellectual equivalent of studying the retreat speed of glaciers to understand sea level rise, while ignoring that the underlying climate system is non-linear.

The real story in their data: 30% of workers have exposure-adoption divergence. That is not a finding about comparative advantage. That is a measurement of implementation lag. That lag will compress. The comparative advantage calculus will stop balancing when one side of the equation stops being variable.

The paper measures the transition's current friction. It does not alter the destination.