Methodology

COOKED assigns every occupation a single headline number, the AI exposure %, plus a second, independent read on how much AI is already in real use. Both trace to named, dated, public sources. Nothing on this site is a hand-picked or invented figure.

The exposure % (capability ceiling)

We reproduce the occupation-level exposure score published in Eloundou et al. 2023, “GPTs are GPTs” (arXiv:2303.10130). Each O*NET task is labelled E0 (no exposure), E1 (direct exposure) or E2 (exposure via an AI-powered tool).

exposure% = round(100 · Σ(coreweight·task_beta)/Σ(coreweight)); task_beta E1=1.0, E2=0.5, E0=0.0; coreweight Core=2, Supplemental=1. Reproduces Eloundou GPT-4 occupation beta (923/923).

In words: task_beta is 1.0 for E1, 0.5 for E2, 0 for E0; each task is weighted 2 if it is core to the role and 1 if supplemental; the exposure % is the weighted mean over the role’s tasks. This exactly reproduces the paper’s published occupation beta for all 923 occupations. We ship the paper’s number, not a bespoke recompute. It is a capability ceiling: what today’s models could attempt, not a forecast of jobs lost.

The second axis (observed real-world use)

observed = Anthropic Economic Index real Claude-usage exposure per SOC (2026); null where AEI did not publish the cell. Source:Anthropic Economic Index (labor market impacts) (CC-BY). The two axes differ on purpose: a high ceiling with low observed use means the capability exists but the workflow hasn’t moved yet.

Bands

Band	Exposure	Occupations
RESILIENT	0–19%	282
LOW	20–34%	164
MODERATE	35–49%	227
HIGH	50–64%	179
SEVERE	65–100%	71

Honest limitations

Eloundou labels are a 2023 capability *ceiling* (what AI could do), not a displacement forecast.
AEI observed exposure grounds the ceiling in real 2026 usage; the two axes differ on purpose.
Exposure measures tasks, not whole jobs; high exposure ≠ your job disappears.
Exposure is measured on the O*NET task list, which lags real jobs and omits the judgment, relationship and accountability work that resists automation.

Sources

O*NET 30.3, U.S. Department of Labor: occupations, task statements, importance.
Eloundou et al., "GPTs are GPTs" (2023): task/occupation exposure labels.
Anthropic Economic Index (labor market impacts): real Claude-usage exposure per SOC.

Dataset generated 2026-07-16 · 923 occupations · 0 join errors.