Every occupation on Jobs After AI carries an AI-Proof Score from 0 to 100. This page explains where that number comes from, which research it rests on, how it was built, and what it cannot tell you.
What the score measures
The AI-Proof Score is a 0–100 estimate of how exposed an occupation's core tasks are to AI over roughly the next three to five years. Higher means safer. A score of 16 means most of the day-to-day work already sits within reach of systems in production today. A score of 94 means very little of it does.
It is an occupation-level baseline: it describes the typical task mix of a job, not any particular person doing it. The quiz starts from that baseline and moves it based on your own answers — what share of your week is routine processing, how digital your output is, whether you already use AI in the work, and how senior you are.
score = clamp( occupation_baseline + 0.65 × Σ(task / AI-use / stage modifiers), 8, 96 )The 0.65 damping factor exists because the baseline already reflects the typical task mix of the role. Without it, the personal modifiers would count the same evidence twice.
What it does not measure
- It is not a probability of job loss. The underlying research measures the exposure of tasks — which sits upstream of, and is softer than, displacement. Anthropic's labour-market work finds no systematic rise in unemployment among exposed workers through early 2026; the clearest signal so far is a tentative slowing of hiring for under-25s in exposed roles.
- It is not a claim about you. Two people with the same job title can have very different exposure depending on what they actually spend the week doing. That gap is what the quiz's modifiers are for.
- It is not a dated forecast. No part of this system predicts that a job disappears in a given year. Where timing matters, claims are tagged by horizon — now, the next few years, or the long term — and the evidence for each is cited separately.
The evidence base
Baselines are built from four occupation-level source families. They measure different things and were published in different years, so they are weighted rather than averaged:
| Source | What it measures | Data vintage | Weight |
|---|---|---|---|
| Microsoft Research — AI applicability score | Real generative-AI overlap with an occupation's work activities, derived from ~100,000 anonymised Bing Copilot conversations | 2024 data, published 2025 | 0.40 |
| Anthropic Economic Index — observed exposure and usage | Revealed AI usage per occupation, the automation/augmentation split, and the gap between capability and actual coverage | 2025–2026 | 0.25 |
| Eloundou et al. — GPT exposure (β) | Theoretical share of tasks an LLM could complete with at least 50% time saving | early 2023 | 0.20 |
| Felten et al. — AI Occupational Exposure | Broad-AI (perception and prediction) exposure, pre-LLM | 2021 | 0.15 |
Two further sources act as scaffolding rather than signal: the O*NET occupational database, which supplies the task substrate all four measures are built on, and the BLS 2024–34 employment projections, used directionally to sanity-check demand. The World Economic Forum's Future of Jobs Report 2025 supplies the fastest-declining and fastest-growing lists.
Why the newer sources carry more weight
Felten (2021) and Eloundou (2023) describe pre-agent capability. Where those indices disagree with 2025–2026 usage data, the newer data decides the current horizon and the older capability ceiling informs the mid- and long-term picture.
The single most important fact behind that ordering: theoretical capability runs far ahead of actual coverage. Computer and mathematical work scores about 94% theoretically exposable on Eloundou's measure, while Anthropic observes roughly a third of it actually covered in real usage today. Capability indices set the ceiling. Usage data says where we currently are. A score built on either one alone would be wrong in a predictable direction.
How a baseline is built
- Anchor. Where an occupation has a direct Microsoft applicability score, or a close analog in the same occupational code, that score is the primary anchor. Where it does not, the baseline is anchored on the nearest scored occupations plus the role's membership in the Eloundou fully-exposed, Felten high-exposure, Anthropic exposed-or-zero, and WEF declining-or-growing lists.
- Invert to a safety scale. Higher exposure produces a lower baseline. The observed exposure scales are mapped onto the 0–100 safety axis so that the most exposed digital roles land in the teens and twenties and zero-coverage physical roles land in the eighties and nineties.
- Apply human-moat adjustments. Usage data systematically under-sees some occupations, because chatbots only touch digital tasks. Four moats add safety on top of the inverted exposure number (below).
- Apply exposure penalties. The inverse case: purely digital output, routine and repetitive tasks, transactional work with no relationship or accountability attached, and presence on the WEF fastest-declining list.
- Calibrate to a wide distribution. Final clamp 8–96, with the spread deliberately preserved rather than compressed toward the middle. A score that returns everyone "55, mixed" tells no one anything.
The four human moats
- Physical embodiment and dexterity — trades, hands-on care, repair. The largest moat, because robotics lags language models badly.
- In-person accountability — work where someone must be present, liable, or trusted: nursing, teaching, therapy, skilled sales.
- Regulation and licensure — law, medicine, audit sign-off, financial advice. These slow deployment even where capability is already high.
- Bodily-safety and duty of care — paramedics, firefighters, childcare. Society resists full automation here regardless of what the technology can do.
Confidence levels
Every occupation carries a confidence level alongside its score, and it changes how firmly the accompanying analysis is written:
- High — a direct applicability score (or a near-identical occupational analog) corroborated by at least one other source family, with no strong cross-source conflict.
- Medium — anchored on nearest-neighbour scores plus tier membership across two or more sources, with some interpolation.
- Low — no direct score, single-source or conflicting evidence, or a judgment-heavy moat call. Pages for these occupations say so in the text, and the score is presented as directional.
Where the sources disagree
Six divergences were resolved by hand rather than by formula. They are listed here because a reader who knows the literature will spot them:
- Software development. Highest current usage exposure of any field — Anthropic observes about 75% task coverage for programmers — yet demand projections are the strongest of any role and the augmentation upside is the largest. Software roles are placed mid-scale rather than at the bottom, split by seniority. This is a judgment call, not a number read off a chart.
- Radiologists. Perception-AI indices and the popular narrative both say high exposure; the working reality is augmentation plus licensure and malpractice liability. Placed moderate.
- Psychologists, counsellors, social workers. Felten's prediction-heavy 2021 index ranks them highly exposed; LLM-era usage is low and the therapeutic-alliance moat is real. The newer sources win; the 2021 ranking is treated as vintage-distorted.
- Accountants, auditors, tax preparers, bookkeepers. A rare four-way agreement across capability, exposure, usage and declining-demand lists. Baselines are low, and lower than our own earlier placeholders.
- Translators and interpreters. Translators rank at the very top of the Microsoft applicability list. Live interpreters keep a real-time, in-person, high-stakes moat. The two are scored far apart despite sharing a language.
- Actuaries, financial examiners, genetic counsellors. Top of Felten's exposure ranking on prediction, but licensure, judgment and in-person counselling pull them back to moderate.
What this score cannot tell you
- Usage data reflects who currently uses AI. Knowledge workers, English speakers, higher earners. It under-measures exposure in occupations whose workers simply have not adopted these tools yet — which is not the same as those occupations being safe.
- Exposure is not displacement. It is the measurable thing upstream of it. Every page on this site holds that line.
- Agentic systems are moving the ground. Baselines built on chat-era usage will need re-grounding as agents move tasks from assisted to delegated. This version is a living baseline, not a verdict.
- The middle carries more interpolation than the ends. Around 140 occupations were grounded against a few dozen directly scored anchors plus the four-source tier structure. The extremes are better evidenced than the middle, and the confidence field says which is which.
From the occupation baseline to your own score
The occupation baseline is the starting point, not the answer. Your task mix, how digital your output is, whether you already work with AI, and your seniority all move the number — in both directions. The quiz applies those modifiers and returns the score with the drivers that produced it.
Take the quiz: How AI-proof is your job? The full ranking of all 159 occupations, with the reasoning behind each one, is in the AI Job Impact Report 2026.
Sources
- Tomlinson, K., Jaffe, S., Wang, W., Counts, S., Suri, S. (2025). Working with AI: Measuring the Applicability of Generative AI to Occupations. Microsoft Research (arXiv:2507.07935).
- Anthropic (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic Economic Index.
- Anthropic (2025). Which Economic Tasks are Performed with AI? Anthropic Economic Index.
- Anthropic (2026). Anthropic Economic Index report: Learning curves.
- Eloundou, T., Manning, S., Mishkin, P., Rock, D. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. OpenAI / University of Pennsylvania (arXiv:2303.10130).
- Felten, E., Raj, M., Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence. Strategic Management Journal 42(12).
- World Economic Forum (2025). The Future of Jobs Report 2025.
- National Center for O*NET Development (2025). O*NET Database.
- U.S. Bureau of Labor Statistics (2025). Employment Projections 2024–2034.
Methodology version 0, last updated 4 August 2026. The research library behind it is versioned; when baselines are re-grounded, this page and the affected pages carry a new date.