Which jobs are safe from AI?

What “safe from AI” actually means here

Every occupation below scores at or near zero on the Anthropic Economic Index's observed AI exposure — the share of real Claude.ai conversations mapped to that occupation's tasks. That is a measurement of the present, not a guarantee about the future. Frey & Osborne's 2013 model, built years before large language models existed, disagrees sharply with several of these rankings, which is itself informative: it shows two different kinds of automation pressure — language-model exposure and physical-robotics exposure — move at different speeds and sometimes in different directions.

With that scoping in place, five mechanisms explain almost every occupation that clusters near zero. They are not mutually exclusive — the strongest cases below usually combine two or three — but each is worth understanding on its own, because each has its own separate expiry condition.

Five mechanisms of resistance

All 19, ranked by observed exposure

Sorted lowest to highest on Anthropic's observed exposure. Notice the spread: most of this list sits at 0%, but Judges sits at 31% — the accountability mechanism protects the ruling, not the document review that precedes it.

“n/a” means that study does not cover this occupation's current SOC code, not that its risk is zero. See method and limitations below.

Robotics is moving the physical half of this boundary

Everything above measures exposure to language models. It says nothing about the separate, faster-moving question of physical automation, and 2026 is the year that question stopped being hypothetical. Schaeffler and the robotics company Humanoid have agreed to deploy an estimated 1,000 to 2,000 humanoid robots across Schaeffler's global manufacturing sites by 2032, with the first units live at two German facilities from December 2026. Amazon and BMW are piloting humanoids in fulfillment centres and on assembly lines today, and Hyundai plans to introduce Boston Dynamics humanoids at its Georgia plant from 2028.

The gap is speed and structure, not willingness to try. One hotel operator running a humanoid pilot reported the robot needing several hours to clean a room a human housekeeper finishes in about 40 minutes. Warehouse trials show humanoids running at roughly 70–85% of human picking speed on trained, repeatable tasks, with failure rates under 5% — genuinely useful, but only once a task has been standardised and drilled.

That is precisely the distinction the physical-dexterity occupations in this piece rely on. Bin-picking in a warehouse is structured and repeatable, which is why warehouse and freight roles are the physical jobs most exposed to near-term robotics, and why they sit outside this list. Rewiring a 1920s house, diagnosing a knock in an engine that has never been on this lift before, or patching a roof in wind are unstructured by nature — every job is a slightly different problem, and that is exactly the case current robotics handles worst. The boundary is real today. It is not fixed.

Licensure protects a role, not a task

A license, board certification, or statutory scope of practice is a legal barrier, not a technical one. It restricts who is allowed to perform or sign off on a task; it says nothing about whether a model could do the analysis underneath it. That gap is why radiology, once treated as an obviously AI-proof specialty because it required a license and years of training, has spent the past several years absorbing AI-assisted image review even as the final diagnostic sign-off has stayed with a licensed physician. The license held. The task mix inside it changed anyway.

The same pattern is visible inside this dataset. Judges post the highest observed AI exposure of any occupation on this list — 31%, roughly triple the next-highest entry — because “read documents on pleadings and motions” and “write decisions on cases” are fundamentally text tasks, even though the license, the oath, and the legal authority to rule stay with a named human judge. Licensure is a genuinely strong shield. It protects the parts of a job that require the license, and nothing else, and legislatures do revisit scope-of-practice rules when the economics get compelling enough.

What actually raises your resistance

• Stack mechanisms, don't rely on one. A single trait — being physical, or being licensed — is the weakest version of resistance. A physician assistant combines licensure, accountability and physical examination; that combination is far more durable than any one piece alone.
• Take the unstructured half of your job seriously. If your work happens in a different environment every time — a different building, a different body, a different fault — that variability is your moat. Standardising your own work to make it faster is also what makes it easier to automate later.
• Own the accountability, not just the output. Being the person whose name is on the decision, the diagnosis, or the ruling is a different job than producing a draft of one. Where you can, move toward the sign-off and away from the first draft.
• Build continuity, not just competence. Relational trust compounds specifically because it is with the same human over time. A rotating cast of equally competent providers is a much weaker version of this mechanism than a long-standing relationship.
• Watch your augmentation share. Across this dataset, occupations where AI use skews toward augmentation rather than automation are the ones where people are still firmly in the loop. A falling augmentation share in your own field is the earliest real warning sign, well before headcount moves.

How durable is this, really?

Treat everything above as a snapshot, not a forecast. Three things can erode it independently of anything a language model does. Demand can fall for reasons that have nothing to do with AI — skilled trades are still cyclical with construction spending and interest rates, and BLS growth projections describe average conditions, not guarantees for any one region or employer. Regulation can change — scope-of-practice rules, telehealth authorisations and AI-assisted diagnostic approvals have all moved in the last few years, and licensure only protects a task for as long as the rule stays in place. And robotics investment is accelerating on its own timeline, entirely separate from the LLM progress this dataset tracks; reporting on the humanoid robotics sector points to shipments scaling from roughly 90,000 units globally in 2026 toward the low millions by 2030, concentrated first in exactly the structured, repeatable tasks discussed above.

None of that is a reason to panic about the occupations on this page. It is a reason not to treat “safe from AI” as a permanent label. The honest version of this page is a photograph of August 2026, not a guarantee.

Why we built this

Planetary Labour builds an autonomous go-to-market engine — software that runs marketing work itself, publishing content and building distribution without a team operating it. We sit squarely on the automation side of this question, which is exactly why we would rather map the mechanisms honestly than sell a reassuring list.

What keeps showing up in our own data is the same pattern that shows up here: the work that goes first is well-specified, text-based and repeatable, and the work that remains is judgement exercised on something unstructured, with a named person accountable for the result. That is the argument in our manifesto and across the rest of our future of work writing.

Method, sources and limitations

Every figure above comes directly from the same dataset behind Will AI Take My Job?, computed at build time from the underlying records rather than typed by hand. Nothing here is averaged, rescaled, or blended across studies.

• Employment and median wage: the BLS Occupational Employment and Wage Statistics May 2025 national estimates, released 15 May 2026.
• Growth projections for electricians: the BLS Occupational Outlook Handbook, which projects roughly 9% employment growth from 2024 to 2034 and about 81,000 openings a year.
• Frey & Osborne, Eloundou et al. and the Anthropic Economic Index: see the full descriptions on Will AI Take My Job?. In short: The Future of Employment, GPTs are GPTs, Anthropic Economic Index, GDPval.
• Humanoid robotics deployment: reporting on the Schaeffler–Humanoid manufacturing agreement, Amazon and BMW pilot programmes, and hospitality-sector trials, via AI News and Tech Times' coverage of IEEE Humanoids 2026.

Known limitations

This dataset is US-only, describes current capability and observed behaviour rather than a forecast, and reflects one AI vendor's traffic (Anthropic's), which skews toward technical and professional users. Occupations with very low usage have small samples, so their splits are indicative rather than precise. The curation above is a deliberate editorial selection of occupations that illustrate each mechanism clearly — it is not an exhaustive list of every low-exposure job in the economy. Use the full tool to look up any specific occupation.

Frequently asked questions