Methodology: where every number comes from

Every number on this site comes from a published dataset you can download yourself. This page says exactly which ones, what each figure means, and — just as important — what it does not mean.

The one rule: exposure is not replacement

An exposure figure measures overlap between a job's task list and what AI systems can do or are observed doing. It says nothing, by itself, about whether the job will shrink. The clearest proof sits inside our own data: software developers carry one of the highest exposure ratings in the GPTs-are-GPTs study, and the Bureau of Labor Statistics simultaneously projects their employment to grow 15% by 2034. Cashiers, by contrast, score only moderately on AI exposure yet face the largest projected job losses of any occupation — driven mainly by self-checkout and online shopping, not by language models. Any site that converts an exposure score directly into a "probability your job will be automated" is inventing a number, and we do not do it.

Anthropic's labour-market research (March 2026) makes the same point from the outcome side: as of its publication, researchers found no systematic rise in unemployment for workers in the most AI-exposed occupations. Exposure is a map of where jobs will change. What employers, licensing bodies, customers and workers do with that change decides the rest.

Source 1 — Anthropic Economic Index: observed exposure

What it is: the share of an occupation's O*NET tasks for which AI use was observed in real, anonymised Claude conversation data, published as a job-level dataset alongside Anthropic's March 2026 paper "Labor market impacts of AI" (Massenkoff & McCrory). Range in the data: 0% to about 75% (computer programmers are the highest). What it captures: what people are actually using AI for at work today. What it misses: use of every other AI tool, and workplaces where AI is used through software that embeds it invisibly. 411 of the 756 occupations in the dataset sit at exactly 0 — for most hands-on jobs, that means "not observed", not "impossible".

Source 2 — GPTs are GPTs: rated exposure (β)

What it is: Eloundou, Manning, Mishkin & Rock (OpenAI / University of Pennsylvania, 2023) had human experts and GPT-4 rate every O*NET occupation's tasks for exposure to large language models. We show the paper's β measure: tasks rated directly exposed count in full, tasks exposed with good tooling count at half. The dataset is published by OpenAI; we show the GPT-4 ratings and print the human ratings next to them on each page. What it captures: capability — what the technology could reach. What it misses: cost, regulation, reliability and whether anyone actually wants the machine to do it. Capability ratings are systematically higher than observed use, which is exactly why we show both.

Source 3 — Microsoft "Working with AI": applicability score

What it is: a 0–1 score per occupation built from 200,000 anonymised Bing Copilot conversations, combining how often people used AI for an occupation's intermediate work activities, how completely it handled them, and the scope of impact (Microsoft Research, 2025; data CC BY 4.0). How to read it: it is a score, not a percentage — across all occupations the median is about 0.14 and the highest (interpreters and translators) is about 0.49. What it misses: specialist, industry-specific AI tools — which is why medical transcriptionists score low here while scoring at the top elsewhere: their transformation runs on medical speech AI, not on general chatbots.

Pay, outlook and education — Bureau of Labor Statistics

Pay, ten-year outlook and typical entry education come from the BLS Employment Projections programme (Table 1.2) and the Occupational Outlook Handbook, 2024–34 projection cycle — public-domain US government data. A median pay figure is the midpoint of an occupation's earnings in 2024, not a forecast and not a starting salary. A projection is a modelled scenario built from stated assumptions, not a promise: BLS revises the whole matrix every year, and this site updates with it. Where the Handbook publishes figures for a wider group than the single occupation on a page (for example its combined profile for software developers, QA analysts and testers), the page says so in plain language and in its Sources list.

Tasks and skills — O*NET

The task lists and skills come from the O*NET Database (version 30.0), the US Department of Labor's occupational database, published under CC BY 4.0. The split of tasks into "where AI can already help" and "what stays human" is our editorial analysis, made task by task against a simple test: does the task primarily move text, numbers or images on a screen (AI-assisted today), or does it require physical presence, a licence, legal accountability, or judgement about a specific person (human)? Reasonable people can move a task across the line; the underlying task list is quoted so you can judge for yourself.

The learn-next roadmaps

Roadmap suggestions start from the occupation's own O*NET skill profile and its officially related occupations, then filter for destinations whose BLS 2024–34 outlook is positive or protected by a licence or physical presence. Every roadmap entry that names another occupation states that occupation's projected change next to it, so the advice is checkable against the same table we used to build it.

How the site is updated

New professions are added weekly, in order of reader demand and source coverage; a page is published only when at least two independent exposure sources plus the BLS figures exist for it. When a source publishes a new edition — BLS revises projections annually, O*NET and the research datasets update on their own cycles — the affected pages are regenerated from the new data and re-dated. If a figure on any page is wrong, the contact page is how to tell us; confirmed corrections are fixed, not argued with.

Licences and attribution

BLS material is US public domain and is used with citation, as BLS requests. O*NET data is used under CC BY 4.0 (US Department of Labor, Employment and Training Administration) — attribution here and on each page; no endorsement is implied. Microsoft's Working with AI data is used under CC BY 4.0 with citation of the paper. The GPTs-are-GPTs dataset is OpenAI's published research data. Anthropic's Economic Index datasets are published by Anthropic for research use and are cited to the March 2026 paper.