We do not call any job title safe from AI, and we will not give you a list. The studies cited here score tasks for exposure, not titles for safety [1][2]. Clerical occupations show the highest exposure [2]. Most occupations also include tasks that need human input [2]. So look at the tasks in your week.
That is a less tidy answer than a top-ten list. It is also the one the evidence supports. The rest of this page explains why, and what you can do with it.
Why can't anyone tell you which jobs are safe from AI?
Because the exposure studies cited here work from tasks. One scores the tasks inside each U.S. occupation [1]. Another predicts scores for the tasks in an international list of occupations [2].
One 2023 study scored occupations by how well their tasks match what large language models can do. It found that around 80% of the U.S. workforce could have at least 10% of their tasks affected. About 19% of workers may see at least half of their tasks affected [1]. The authors state that they make no predictions about how fast any of this happens [1].
Read those two numbers together. About four in five U.S. workers could see at least a tenth of their tasks affected. About one in five could see half or more affected [1]. That leaves about one in five below the first mark. The study also lists 34 occupations in which it labeled no tasks as exposed [1]. So exposure is wide and uneven, but it is not universal. A label of no exposure is one rubric's reading, not a promise.
The International Labour Organization measured something different in 2025. It looked at workers worldwide and counted any exposure in an occupation. By that measure, one in four workers are in an occupation with some exposure to generative AI [2]. Only 3.3% of global employment falls in its highest exposure category [2]. The two studies use different groups and cutoffs, so do not compare their figures directly. The ILO's conclusion is worth quoting: "As most occupations consist of tasks that require human input, transformation of jobs is the most likely impact" [2].
So a list of safe job titles answers a question these studies do not ask. In our view, a title hides the mix of tasks under it. Do not assume your week matches the week of someone else with your title.
Which kinds of tasks does AI reach first?
Start with an older argument. In 2001, three economists studied how computers change the skills that jobs demand [3]. They argued that computers substitute for "routine (repetitive) cognitive and manual tasks" [3]. Computers, they wrote, execute rules-based logic [3]. Their data ran from 1960 to 1998. It linked computerization to falling relative demand for routine tasks [3].
Newer sources study generative AI. The ILO points to its "expanding abilities" on specialized tasks [2]. Here is what those newer sources report:
- Clerical tasks. The ILO reports that clerical occupations continue to have the highest exposure levels [2].
- Tasks done fully on a screen. The ILO also found that some strongly digitized occupations now show increased exposure [2].
- Simple, repeated requests. A federal review of banking chatbots found they may be useful for resolving basic inquiries [4].
The Bureau of Labor Statistics puts it in one line. It expects AI to mainly affect occupations "whose core tasks can be most easily replicated by GenAI in its current form" [5].
Notice the word "core." If the exposed tasks are a small part of your week, the picture differs. If they are most of your week, it differs again.
Which kinds of tasks hold up, and why?
The same 2001 paper argued the other half of the pattern. In its words, computers complement "activities involving non-routine problem solving and interactive tasks" [3]. We read "complement" to mean the tool adds to what the person can do.
That gives you two kinds of tasks to look for.
Problems with no script. These are cases your process did not plan for. The federal chatbot review saw this limit in practice. It found that chatbot effectiveness "wanes as problems become more complex" [4].
Work that happens between people. Calming a customer down is one example. So is getting two teams to agree, or reading what your manager meant. Those examples are ours. The 2001 paper's term is interactive tasks [3].
We add a third question in our kits: what happens if the task goes wrong? This is our method, not a research finding. When a mistake is costly, someone has to check the work and answer for it. A tool can draft a refund decision. A person still has to stand behind it.
None of these task types is a guarantee. The 2001 paper is about computers, not today's AI tools [3]. The chatbot review dates from 2023 [4]. Treat both as a starting point for your own check, not as a verdict.
Does higher pay or more education protect a job?
Not by itself. The language-model study found that the projected effects span all wage levels. Higher-income jobs may face greater exposure, not less [1]. Its analysis also suggests that people with bachelor's, master's and professional degrees are more exposed than people without formal credentials [1].
The ILO saw something close to that. Exposure is growing in professional and technical roles as the tools get better at specialized tasks [2].
Exposure is not the same as job loss, though. As we use the word, it means a tool can reach the task. It does not tell you what an employer will do next. The Bureau of Labor Statistics says the timing and scale of many potential impacts are too uncertain to be reflected in its projections [5]. It also notes that such changes tend to occur gradually [5].
So do not count on a degree or a salary to settle the question. Look at your task mix.
Are there any AI-proof jobs?
We do not call any job title safe from AI, and we will not give you a list. Most occupations also include tasks that need human input [2].
What jobs will AI not replace?
We cannot tell you that, and we would doubt anyone who says they can. The exposure studies above measure which tasks a tool could reach, not which jobs end [1][2]. The first one makes no predictions about adoption timing [1]. The Bureau of Labor Statistics does publish forecasts. It says projecting future employment "involves substantial uncertainty" [5].
What you can know is narrower and more useful. You can know which of your tasks a tool already does well. You can know which ones still need you, and why. You can know where your time goes each week.
That is a question about your job, not about jobs in general. You are the only one who can answer it.
How do you check your own job?
List your tasks for one normal week. Be specific. "Email" is not a task. "Answer vendor questions about late invoices" is.
Then rate each task on three things:
- AI reach. Can a tool do this now, with the access it would really have?
- Stakes. What does a wrong answer cost, and who pays for it?
- Human edge. Does this task lean on judgment, trust or context that you hold?
Each task then gets a call. Hand it off, pair with a tool, own it, or cut it. You can read the full steps on our page about how a task audit works.
Next, practice the skill that sits under all of this. That skill is checking an AI's work before anyone acts on it. Our free practice case gives you a simulated AI analysis with twelve findings to judge. Checkout asks only for your email.
If you want the full method for your role, we publish career kits for nine office roles. Each one covers one job. None of them promises a job, a raise or a promotion. They help you see your work clearly and show what you did with it.
What else do people ask about jobs and AI?
Which jobs are safe from AI?
We do not call any job safe. The studies we cite score tasks for exposure, not job titles for safety [1][2]. Most occupations include tasks that require human input [2]. Look at the mix in your own week.
What jobs will be safe from AI in the future?
We will not guess. The authors of one exposure study make no predictions about timing [1]. The Bureau of Labor Statistics calls the timing and scale of many impacts too uncertain to project [5].
Are high-paying jobs less exposed to AI?
No. One major study found projected effects across all wage levels. It found higher-income jobs may face greater exposure to language models [1]. Pay does not tell you which of your tasks a tool can reach.
Does AI exposure mean my job will be cut?
Not on its own. As we use the word, exposure means a tool can reach a task. The ILO says transformation of jobs is the most likely impact, because most occupations include tasks that require human input [2].
Sources
- "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models," Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock, arXiv (Cornell University), 2023. https://arxiv.org/abs/2303.10130
- "Generative AI and Jobs: A Refined Global Index of Occupational Exposure," ILO Working Paper 140, International Labour Organization, 2025. https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure
- "The Skill Content of Recent Technological Change: An Empirical Exploration," David H. Autor, Frank Levy and Richard J. Murnane, NBER Working Paper 8337, National Bureau of Economic Research, 2001. https://www.nber.org/papers/w8337
- "Chatbots in consumer finance," Consumer Financial Protection Bureau, 2023. https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/
- "Incorporating AI impacts in BLS employment projections: occupational case studies," Christine Machovec, Michael J. Rieley and Emily Rolen, Monthly Labor Review, U.S. Bureau of Labor Statistics, 2025. https://www.bls.gov/opub/mlr/2025/article/incorporating-ai-impacts-in-bls-employment-projections.htm