Observed Exposure (Massenkoff–McCrory Measure)

Observed Exposure

One-line

An occupation-level AI displacement-risk measure that weights what LLMs can theoretically do by how they are actually being used in work, weighting automated > augmentative uses. Introduced by massenkoff-mccrory-labor-market-impacts-2026.

Construction

Three inputs combine at task level via O*NET (~800 US occupations × ~20k tasks):

1. Eloundou β (eloundou-beta-exposure) — theoretical LLM capability score per task: 1 (LLM alone halves task time), 0.5 (LLM + tools), 0 (neither).
2. Anthropic Economic Index usage — share of observed work-related Claude conversations matched to each O*NET task; whether usage is automated (full weight) or augmentative (half weight).
3. Occupation employment weights — task-level scores averaged to occupation via O*NET time-fractions, then to category via BLS employment.

Output: an occupation-level scalar between 0 and 1 representing the share of an occupation’s work-time AI is observed to cover.

What makes it different

Prior task-exposure measures (Eloundou et al. 2023; Acemoglu et al. 2022; Felten–Raj–Seamans 2021) project displacement risk from capability alone. Observed Exposure adds a realized-deployment layer: a task is only counted if it has sufficient real Claude traffic in professional settings.

The empirical payoff: BLS 2024–2034 occupational growth is anti-correlated with this measure (slope −6.07, R² 0.027) but not with Eloundou’s β alone. Realized usage adds predictive value the theoretical measure lacks. See ai-exposure-vs-bls-growth.

What it reveals

• A large gap between feasible and deployed. Computer & Math: 94% theoretical / 33% observed. See theoretical-vs-observed-capability-gap.
• Who the most exposed workers are. Computer programmers (74.5%), customer service reps (70.1%), data entry keyers (67.1%). See most-exposed-occupations.
• 30% of workers with zero exposure — cooks, mechanics, lifeguards, bartenders, dishwashers, dressing room attendants — the manual-physical floor.

What it does not capture

• Forward-looking shifts. The measure is anchored in current Claude product surfaces (mid-2025 usage data). If frontier capabilities reach previously inaccessible task categories, the measure lags reality.
• Cross-vendor diffusion. Claude usage ≠ all LLM usage. ChatGPT, Copilot, Gemini, open-weight deployments are not captured.
• Augmentation surplus. Augmentative use is half-weighted; this is a methodological choice with consequences — augmentation that meaningfully raises productivity (without reducing worker count) is counted as half a displacement signal.
• General-equilibrium effects. Demand reallocation, new task creation, wage adjustment are downstream of the measure, not in it.
• Conflict of interest. The vendor measuring its own labor-market impact is itself the deployment surface. The published HuggingFace dataset is reproducible but does not resolve the structural conflict.

Relation to other exposure frameworks

• Eloundou β (2023): theoretical only. This measure uses β as input but adds usage.
• Acemoglu et al. (2022): uses online vacancy data. Industry-level rather than occupation-level. Different mechanism (firm posting).
• Hampole et al. (2025): university-hiring-network IV approach. Firm-level adoption causally identified. Different unit of analysis.
• Brynjolfsson et al. (2025): ADP payroll panel. Age-group focused. Compatible with this measure on the young-worker finding (see ai-young-worker-hiring-slowdown).
• Tomlinson et al. (2025) “applicability”: measures the type of task–LLM fit. Adjacent rather than competing.

Why it matters to Where Value Lands

• Provides the first labor-side measure that closes the loop between capability and deployment. Capability without diffusion is invisible to the labor market; diffusion without capability is impossible. Observed Exposure is the intersection.
• Operationalises diffusion-vs-innovation at occupation granularity.
• Anchors the labor side of the H2_u-curve-of-value story: exposure clusters at high-skill knowledge work, not low-skill service work.
• Establishes a detectability threshold (~1pp differential unemployment) for any future displacement claim.

Related

• massenkoff-mccrory-labor-market-impacts-2026 — source.
• theoretical-vs-observed-capability-gap — the headline finding the measure exposes.
• exposed-worker-demographics — who scores high on this measure.
• ai-young-worker-hiring-slowdown — first empirical signal from applying the measure.
• most-exposed-occupations — top-10 list under this measure.
• ai-exposure-vs-bls-growth — predictive validation against independent BLS forecasts.
• diffusion-vs-innovation — the broader frame this measure operationalises.
• task-based-framework — Acemoglu’s macro frame this measure sits inside.
• capital-labor-divergence — labor-incidence story this measure feeds.
• scaling-gap — capability-vs-deployment gap at the enterprise layer; parallel phenomenon.
• autonomy-slider — the automated-vs-augmentative weighting echoes the autonomy spectrum.
• fusion-skills — the augmentative side of the weighting; what gets discounted at half weight.
• displacement — the concept this measure is the empirical instrument for.