Methodology

The AI Task Exposure Score is deterministic, versioned, explainable, and reproducible. This page is the complete definition — there is no hidden model behind the number.

What the score is

Each occupation is decomposed into its constituent tasks. Each task carries a structured exposure value from 0 (AI contributes little) to 100 (current AI systems can perform most of the task), and a classification: automatable, augmentable, or durable. The occupation score is the share-weighted aggregate:

occupation_exposure = Σ(task_share × task_exposure) / Σ(task_share)

In the current release (2026.07.12-r2), task shares are equal within each occupation — the corpus does not yet include per-task time weights. That is a documented limitation, not a hidden assumption; share calibration is on the roadmap, and a share change will create a new score version. Every occupation page shows its own arithmetic in a “How this score was calculated” box.

How each task’s exposure value is scored

Every one of the corpus’s 16,578 task values was scored against the anchored rubric below (data release 2026.07.12-r2), one occupation at a time so values are calibrated within the role, using a frontier language model (claude-sonnet-5) as the scoring instrument. The anchors define the intended scale and the released values remain structured estimates. Once published, title results are calculated deterministically from that fixed release without a model call. The complete before/after audit trail is retained.

EXPOSURE SCALE (anchored bands):
 0–15  Physical work in unpredictable real-world environments; hands-on care,
       force, dexterity, or physical presence is the essence. No current
       AI/robotic system performs this in the field.
16–35  In-person interpersonal or physical core with meaningful digital
       periphery: face-to-face persuasion/authority/comfort, embodied
       inspection, on-site coordination. AI can prep or summarize around it.
36–55  Genuinely mixed: the human decides/executes (often in person or with
       legal accountability) but AI already does substantial parts — drafting,
       triage, analysis, planning — under supervision.
56–75  Predominantly digital/cognitive routine: analysis, documentation,
       scheduling, standard communication, coding, research. Current AI does
       the bulk well with human review.
76–100 Fully digital and rule/pattern based end-to-end: data entry,
       transcription, routine classification, template documents. Current
       AI/software performs it with minimal oversight.

CLASSIFICATION:
durable      physical presence/manipulation, embodied care, high-stakes
             in-person judgment, legal authority, or accountability is the core
augmentable  the human owns the task; AI materially accelerates parts of it
automatable  AI/software can perform it end-to-end with quality control

Data release history — including our corrections

2026.07.11-r1 — initial import of the occupation/task corpus. Its task values came from a keyword-based classifier that systematically over-scored physical and in-person work (example: “render aid to accident survivors” was scored 44/100, and 46% of occupations had no task classified durable at all).
2026.07.12-r2 — every task exposure value and classification re-scored against the anchored rubric above. Physical, protective, and care occupations moved sharply down (patrol officers 55 → low 20s); routine digital work stayed high. The scoring formula did not change — only the underlying task data. Read the full correction story.

What the score is not

This score estimates how exposed the tasks in a role are to current and near-term AI capabilities. It does not predict whether a specific person will lose a job. It is not a probability of firing, an “AI replaces you by 20XX” date, or an actuarial forecast. We never publish those claims, and the paid report is prohibited — by validation code, not just policy — from inventing them.

Percentile, rank, and bands

Percentile = share of the 968-occupation corpus with a strictly lower score. Tied scores share the same percentile — equal scores are never displayed as different.
Rank uses competition ranking (1 = most exposed); tied scores share the rank.
Bands are corpus quartiles, not emotional thresholds: Low ≤ 23, Moderate ≤ 34, High ≤ 43, Very High above. Labels stay stable relative to the analyzed labor market.

The personalized score (paid report)

The $9 report recalculates the same formula over your task mix: an AI model extracts and normalizes tasks from the text you paste, proposes taxonomy mappings, and estimates time shares when you did not provide them. The application then resolves and validates IDs, applies the numeric formula, and computes percentile, band, and confidence deterministically; the model does not calculate the final score. Activities that don’t match your occupation’s taxonomy get a second matching pass against closely related occupations’ tasks (at reduced mapping weight); anything still unmatched is listed but excluded from the number — it reduces the stated confidence instead of silently distorting the result. The extraction prompt tells the model to ignore cover-letter narrative, objectives, and personal qualities that are not work activities.

Confidence

Reported as Low / Medium / High with the reason, computed from observable coverage, mapping confidence, related-taxonomy matches, input length and quality, and whether task shares were supplied or inferred. We do not fabricate decimal certainty.

Lower-exposure adjacent roles

A role is labeled lower-exposure only when all of the following hold under the same score version: it scores at least 10 points lower, skill overlap is at least 50%, and both roles have current score data. In the current release only 5.3% of the adjacency graph qualifies — the rest are shown as “related roles” with no safety claim. We consider that honesty a feature.

Data provenance and versioning

Occupation, alias, task, and skill data originate from a versioned corpus derived from public occupational data (O*NET-linked task statements with structured exposure values), imported once as data release 2026.07.12-r2 with a recorded SHA-256 checksum.
Scores are recomputed from raw tasks on our side — release jr-v1, published 2026-07-12.
Every stored result records its data version, score version, and (for paid reports) prompt and model versions. A methodology change creates a new score version; historical reports keep their numbers.
Labor-market context (median wage, employment, openings, projected growth) is display context only — it is never an input to the exposure score.

Sources, license, and independence

This product includes occupational information derived from the O*NET database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under the CC BY 4.0 license. JobAIRisk has modified and combined some source information, including its task-exposure estimates; USDOL/ETA has not approved, endorsed, or tested those modifications. O*NET® is a trademark of USDOL/ETA.

Displayed U.S. employment, openings, wage, and projection fields are contextual occupational data from the imported source snapshot and are aligned to the U.S. Bureau of Labor Statistics Employment Projections occupational data. They are not inputs to the exposure score. JobAIRisk is independent of USDOL/ETA, O*NET, BLS, Anthropic, Google, Microsoft, and OpenAI.

Known limitations

Equal task weights within an occupation (until share calibration ships).
Task exposure values are structured estimates of current and near-term AI capability against typical task content; they are periodically revised via new data releases, not live-updated.
Occupation-level results average over real variation between employers, seniority levels, and regions — that is exactly what the personalized report exists to correct.
The adjacency graph measures skill overlap, not hiring demand; a lower-exposure adjacent role is not a guaranteed or recommended career outcome.

Questions about the method? Contact us — corrections are published on this page.