Any useful answer starts by explaining why the two standard methods give opposite rankings, because almost every published list picks one method, does not say which, and presents the result as fact.
The answer, in one line
There is no defensible ranked list, because the two standard methods give opposite answers. Task exposure asks what proportion of a job's tasks a model could perform, and ranks well-paid desk-based work as most exposed.
Two methods, two answers#
Task exposure asks what proportion of a job's tasks a model could perform. It is the method behind the founding statistic of the field: around 80 per cent of US workers could have at least 10 per cent of tasks affected, and about 19 per cent could see at least half affected. On this method, the most exposed jobs are well-paid, educated, desk-based, and the safest jobs are physical.
Substitution versus complementarity asks something different: when AI touches this work, does it replace the person or make them more effective? Stanford's payroll analysis found declines concentrated in occupations where AI substitutes for human tasks, while employment was flat or rising where it complements, especially for experienced workers.
These rank differently because exposure measures what is technically possible and substitution measures what organisations actually do. A radiologist is highly exposed and, so far, complemented. A junior copywriter may be less exposed on paper and more substituted in practice.
Four things that predict safety better than occupation#
The occupational frame is the problem. These cut across job titles.
1 · Accountability that cannot transfer. Where someone must be answerable, and a system cannot be, the human role survives even when the machine is more accurate. This is why regulated professions are more durable than their task exposure suggests, and Article 14 of the EU AI Act has now made some of it law.
2 · Context the model cannot have. Work depending on knowing this organisation, this client, this history. Most senior work is largely this, which is a better explanation of seniority's durability than the tasks themselves.
3 · Physical presence with variation. Not physical work generally, which robotics is reaching, but physical work in unpredictable environments. A plumber in an unfamiliar house is doing something genuinely hard to automate.
4 · Relationships where being human is the point. Care with a hard edge: AI-generated replies have been rated as making recipients feel more heard than replies from untrained humans, and labelling the reply as AI removed the advantage. What is valued is that a person chose to attend to you, which is not a performance property.
The exposure nobody lists#
The clearest risk is not an occupation at all. It is a position in a career. Stanford found the effect concentrated in 22 to 25 year olds in exposed occupations, running through reduced hiring rather than redundancies. Two people in the same job title, one with fifteen years of context and one with none, face completely different exposure. Every list organised by occupation misses this. It is the largest single finding in the current evidence base. Occupation is the wrong unit.
Exposure is not adoption#
Considerable. Exposure studies measure capability rather than adoption, and adoption is slower and stranger than exposure predicts. Stanford's finding is observational and cannot establish causation. Aggregate labour-market effects remain small: Danish evidence found precise null effects on earnings and hours two years after ChatGPT, ruling out effects larger than 2 per cent. And Acemoglu's modelling puts total factor productivity gains at under 0.66 per cent over a decade, which is not the profile of a technology reorganising the labour market quickly.
Anyone publishing a confident ranking of safe jobs is working from exposure estimates and presenting them as forecasts.
A question about hiding#
"Which jobs are safest" is a question about hiding. It assumes a stable place exists and the task is to find it, which is a reasonable thing to want and a poor description of how this is unfolding.
The better question is which capabilities appreciate, because those move with you and no occupational category protects you if you lack them. Judgement in a domain, the ability to detect when a confident answer is wrong, accountability others will accept, and context that took years to accumulate. Those are not safe. They are valuable, which is a stronger position than safety and the only one actually available.
Related SuperSkills research#
On the individual question, will AI replace my job. On career position, entry-level jobs and the missing rungs. On what appreciates, staying valuable and what stays human.
Key sources
- Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine?
- Eloundou, T. et al. (2023). GPTs are GPTs.
- Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents.
- Acemoglu, D. (2024). The Simple Macroeconomics of AI.
- Yin, Y. et al. (2024). AI-generated replies and feeling heard.
About this research#
Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This page deliberately declines to publish a ranked list, because the available methods do not support one.
Evidence review · SS-2026-089 · Graded against the published rubric
Hirji, R. (2026). Which jobs are safest from AI?. The SuperSkills evidence base, SS-2026-089. https://thesuperskills.com/research/which-jobs-are-safest-from-ai. Last reviewed 26 August 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work