StableJob
Methodology · last revised 8 August 2026

What we measure, what we borrow, and what nobody can currently know.

Two signals. We keep them separate on purpose, we name the source of each, and we list what each one cannot tell you. Every number on this site traces back to something on this page.

0

datasets exist that measure actual job replacement. Ours included.

Structural exposure describes the shape of a job's tasks. Real-world usage describes how much AI is being applied to work like it. Both are leading indicators of work changing. Neither is evidence that a person has been replaced, and no public dataset currently attributes job losses to AI at occupation level.

We could produce a replacement probability. It would be a guess wearing a decimal point. We would rather lose that comparison than pretend to know what nobody can currently know.

Signal 01 · theoretical
Live

Structural Exposure Score

How exposed a job's tasks are to large language models, based on task content and structure alone — not observed usage. We combine two independent research measures: OpenAI's GPTs are GPTs ratings of how much of each occupation's O*NET tasks a language model can plausibly perform, and the AI Occupational Exposure (AIOE) index, a separately-built crosswalk of AI capabilities to occupational abilities. Where both cover an occupation we take a coverage-aware average; where only one does, the result is marked lower-confidence.

Rollout status · 8 August 2026

We've recomputed this score for all 70 jobs using the method described here. We're rolling the new number onto each job page only after that job clears a human-researched checklist — checking licensing rules and physical-presence requirements the model above can't see. A job page showing a Research Verdict panel is displaying the new score; one that doesn't yet is still showing our prior, less rigorous estimate.

Inputs
OpenAI's GPTs-are-GPTs task-exposure ratings (Eloundou et al., 2024) and the AIOE index (Felten, Raj & Seamans, 2021), both mapped to O*NET-SOC occupation codes.
Output
A number, 0–100, plus a confidence level reflecting how many of the two sources actually cover that occupation.
Where it misleads
It assumes capability equals adoption, and it has no concept of licensing walls or physical presence — see "Physical & regulatory blind spot" below for the 7 jobs this affects most. A job can also look highly exposed for a decade and change very little in practice.
Signal 02 · measured
Live

Real-World AI Usage

How much AI is actually being used on this job's tasks today. Our primary source is Microsoft's Working with AI study, which maps observed assistant conversations onto occupational work activities. We align those activities to the same O*NET spine the structural score uses, then report a band — corroborated, where it has genuine coverage, by the Anthropic Economic Index.

Input
Microsoft Research, Working with AI: measuring the occupational implications of generative AI (2024) — the primary source for every band. The Anthropic Economic Index (Claude.ai/API usage) adds an independent corroborating note on the 22 of 62 jobs where it has real signal.
Output
Low, Medium or High. Never a decimal — the underlying data does not support that precision and its authors say so plainly.
Where it misleads
It only sees work done at a keyboard, through one product, in one year. Absence of measured usage is not evidence of safety — it is often evidence that the work is not measurable this way at all.

Structural score sources

Two independent research measures, named directly rather than left as a vague "algorithm."

OpenAI — "GPTs are GPTs"
Primary structural input · live
Eloundou, Manning, Mishkin & Rock, Science 384 (2024). Human- and GPT-4-rated exposure of O*NET task statements at the occupation level. We use the "with software tools" (β) gradient rather than the raw or maximal gradients the paper also publishes.
Coverage: 67 of 70 jobs.
AIOE — AI Occupational Exposure
Primary structural input · live
Felten, Raj & Seamans, Strategic Management Journal 42(12) (2021). An independently-built crosswalk of AI capability progress to the occupational abilities O*NET tracks — not an LLM rating, a different method entirely from the OpenAI source above.
Coverage: 66 of 70 jobs.

How the two combine

We never average them. Each job is placed on both axes, and the position is the finding. Four honest categories, no ranking.

Assisted in practice
Changing now
Insulated today
Exposed on paper only
Structural exposure →
Insulated today
Low on paper, low in measured usage. Frequently physical work — which is also where our measurement is weakest, so read the caveat on the page.
Exposed on paper only
Theory says automatable; measured usage has not arrived. Worth watching, not worth panicking about.
Assisted in practice
Heavy usage on a job the model calls hard to automate. AI is being used as a tool by the people doing the work.
Changing now
Both signals elevated. The clearest case that the work is being restructured — and still not evidence that jobs are being removed.

Limitations, per source

Each of these weakens a specific claim we make. They are here so you can discount our numbers exactly where they deserve it.

Microsoft usage data
Primary usage source · live
A static 2024 snapshot of one assistant's usage. It does not update as adoption moves, and it reflects that product's user base rather than the whole workforce.
Weakens: any claim about usage levels today, and about workplaces that do not use this tooling at all.
Keyboard bias
Structural to the method
Usage can only be observed where people type. Manual and trade occupations therefore show low bands partly because their work is invisible to this kind of measurement.
Weakens: the reassuring low bands on trade pages. We repeat this caveat on each of them rather than only here.
O*NET task data
Structural input · live
Task descriptions lag real practice by several years and describe an average version of a job that few people hold exactly.
Weakens: structural scores for fast-changing, hybrid or highly specialised roles.
Coverage gaps
Editorial rule
Many occupations do not appear in the usage source at all. Where that happens we publish no usage band rather than borrowing one from a neighbouring job.
Weakens: comparability across the index. A blank column is a real answer, not an omission.
Physical & regulatory blind spot
Known gap · disclosed
Both structural-score sources rate task content — neither has any concept of who is legally permitted to do a task or whether it requires a body in a physical place. For a small set of occupations — CDL Truck Driver, Bus Driver, Heavy Equipment Operator, CNC Machinist, Quality Control Inspector, Elevator Installer & Repairer, and Groundskeeper / Grounds Maintenance Worker — this produces a structural score materially out of step with what the job actually requires. We flag this directly on each of those job pages rather than only here.
Weakens: the structural score specifically on these 7 jobs. Read the flagged note on their pages before trusting the number.
Anthropic Economic Index
Corroborating source · live
A second, independent usage signal (Claude.ai/API conversations) shown only where it has genuine non-zero coverage — 22 of our 62 jobs today, almost entirely white-collar or licensed roles. We do not have a reliable, per-occupation source for its automation-versus-augmentation split, so we do not publish that shape yet.
Weakens: any expectation of full coverage. Most trade and manual occupations show no Claude signal at all, which the index's own authors attribute to the product being keyboard-bound, not to those jobs being AI-proof.
Census Bureau BTOS
Industry-level corroboration · live
Firm-reported AI adoption by NAICS sector, shown on industry pages only (8 of our 9 industries — Emergency Services & Public Safety has no private-sector NAICS match, so we publish none rather than force one). It measures firms, not jobs.
Weakens: any attempt to read this as a job-level number. It never feeds a structural or usage score — it is context for the industry as a whole, nothing finer.
Not currently used

Sources we are evaluating for future coverage.

None of these feeds any score on this site today. We list them because you may have seen them cited elsewhere, and because we would rather name our roadmap than let anyone assume it is already built.

WEF Future of Jobs
Employer expectations rather than measurement. Useful as narrative context; we have not found a defensible way to turn stated intent into a score.
Evaluating
Sources in use
Microsoft Research
Working with AI: measuring the occupational implications of generative AI, 2024. Primary usage signal.
O*NET / US DOL
Task statements and work-activity weights behind the structural score.
BLS
Occupational definitions, median pay and projected growth. Not part of either signal.

StableJob is an information resource. Nothing here is career, financial or investment advice, and no outcome is guaranteed for any individual. Scores are recomputed when a source publishes an update; every job page carries its own last-updated date.

Browse careers →Browse industries →