Who Uses AI? Platforms, Workforce, and AI Exposure

THE DISSECTION

This paper performs methodological surgery on the empirical AI-labor literature. The finding: current occupational AI exposure scores are contaminated artifacts of platform user demographics rather than genuine occupational vulnerability. The numbers everyone cites—Autor, Acemoglu, Webb, et al.—are measuring who uses AI platforms, not who AI can replace.

THE CORE FALLACY

The entire "AI exposure" literature operates on a category error: conflating AI adoption with AI capability. A platform's conversation logs encode which demographics chose to use that platform, not which occupations are structurally substitutable by AI. When undergraduate humanities majors heavily use ChatGPT, the score says "文学" is AI-exposed. When AI-literate workers in finance use enterprise tools, the score says finance is exposed. These are sampling artifacts, not economic facts.

The authors demonstrate this cleanly: same outcome variable, same controls, different AI platform → employment coefficient changes by 1.9x. Same vendor, consumer vs. enterprise channel → estimates disagree in sign. The entire literature is measuring noise shaped like demographics.

HIDDEN ASSUMPTIONS

1. Platform conversation logs are a random or representative sample of work tasks—which they are not
2. Current AI adoption rates proxy AI capability ceiling—which they don't
3. BLS occupational employment shares are stable enough to reweight toward—which they aren't (the paper corrects for this and the signal attenuates massively)
4. The measurement error is classical (classical measurement error attenuates estimates toward zero)—it is not; it is non-classical, which means the bias direction is unknown and the paper explicitly shows it understates substitution more than augmentation

That last point is the knife in the ribcage of the optimistic reading.

SOCIAL FUNCTION

This is partial-truth self-correction from within the academy—the epistemic community acknowledging their instruments are broken. It is not copium, not propaganda. It is methodological honesty that will be weaponized by all sides: "See, we don't know the effects" (optimist reading) and "The true effects are worse than measured" (realist reading).

The paper's own conclusion undermines the optimistic reading: the bias understates substitution more than augmentation. Correct for the measurement error, and the signal gets worse, not better.

THE VERDICT

Under the Discontinuity Thesis framework, this paper is forensic confirmation of a specific mechanism: the empirical literature is systematically undercounting AI substitution effects because its measurement instruments conflate adopter demographics with occupational structure. The lag-weighted collapse is real; the measurement tools are too broken to see it clearly.

This paper is a better argument for rapid structural change than most of the DT literature—precisely because it arrives from within mainstream economics and still lands on "our instruments are broken in the direction of understatement."

The calibration is wrong. The direction is not in doubt.