Are there jobs AI will never replace? The evidence by task

01Why 'never' is the wrong word

In 2013, Oxford researchers Carl Frey and Michael Osborne estimated that about 47 percent of US employment was at high risk of computerisation. They also judged that tasks needing creativity or social intelligence were unlikely to be taken over by computers within a decade or two.

In 2026, the US Bureau of Labor Statistics (BLS) places writers, editors and graphic designers in its Very high AI exposure group. The two studies measure different things, so this is not a like-for-like test. It does show how quickly a 'safe' list can age. BLS also says a Low label does not mean an occupation is safe from technological change.

02Five kinds of work the evidence treats as harder for AI today

These are task types, not job titles. Each comes with the evidence and where it runs out.

• Hands-on work in changing physical settings. The BLS Low group is full of it: carpenters, construction laborers, maintenance workers and power-line workers. Limit: the studies behind the groups mostly cover language models, so they do not test physical robots.
• Hands-on personal care. Nursing assistants, licensed practical nurses, dental hygienists and massage therapists are all in the Low group in the BLS table. Limit: Low is a relative rank and does not predict job growth.
• Work where a person must answer for the result. The American Bar Association's 2024 ethics opinion keeps lawyers responsible for checking AI output before relying on it. Limit: this protects the sign-off role and says nothing about how many people are hired to draft beneath it.
• Trust between people. BLS's case studies note that robo-advisors compete mainly with younger clients, and that older clients who prefer human advice should keep demand for personal financial advisors strong. Limit: that is a judgement based on information from June 2024, not a measurement.
• Building and powering AI itself. BLS says AI adoption and data centre demand are expected to lift employment in electricity and related infrastructure. Limit: it depends on that build-out continuing.

03Even in 'safe' jobs, many tasks are exposed

Almost no job is untouched. Researchers estimated in 2023 that about 80 percent of US workers could have at least 10 percent of their tasks affected by large language models, and about 19 percent could have at least half affected. The authors made no prediction about timing.

A late-2025 research project called the Iceberg Index estimated that AI systems can already perform skills worth 11.7 percent of US wage value. The authors say the index measures technical exposure, not job losses or when adoption will happen.

04The evidence cuts both ways

Against any safe list: a Stanford team using payroll data finds that employment of workers aged 22 to 25 in AI-exposed occupations is 19 percent below where it would be had they tracked less-exposed jobs, mainly through less hiring. Experienced workers show no comparable gap, and the authors call the results descriptive, not causal.

Against doom: by our count of the BLS table, the Very high exposure group as a whole is projected to grow about 2 percent by 2035, even though 60 of its 206 occupations are projected to shrink. Exposure and job loss are not the same thing.

05How to read any 'AI-proof jobs' list

Use these checks before you trust a list, including ours.

• Check the date. BLS notes that its theory-based sources reflect AI as of mid-2023.
• Ask whether it ranks tasks or job titles. A title hides the mix of tasks inside the job.
• Ask what is measured: exposure, jobs lost or job growth. They are different things.
• Look for a named method and source. Skip a list that gives percentages with none.
• Treat it as a prompt for questions, not a promise.