Will AI replace your job?

No, not on the evidence so far — and the honest version of that is narrower than it sounds. Anthropic measured its own usage against labour statistics in March 2026 and found "no systematic increase in unemployment for highly exposed workers since late 2022". Alongside that: hiring of younger workers in those occupations has slowed. Read the paper.

That is one company's data about one product over three years. It is also the closest thing to a measurement of jobs anyone has published. Nearly everything else measures tasks, and tasks are what are moving now.

So: will your job be replaced this year? Probably not. Will a large share of the tasks inside it end up inside a workflow that calls a model? Probably, yes. The second question is the one you can act on.

Ten measurements, each with its source and its date. They disagree with each other, which is part of the picture.

The useful finding is not that AI replaces jobs. It is that AI changes which part of a job is worth paying for — and it pushes two jobs opposite ways in the same week.

Anthropic's January 2026 index asked what is left for the person once the model has taken the tasks it does well. Travel agents come out deskilled: complex planning is the part a model can now do, leaving routine ticket buying. Property managers come out upskilled: bookkeeping is what moves, leaving contract negotiation and stakeholder work. Read the report.

So "will AI replace my job" is being asked at the wrong altitude. For a job title, the answer is a forecast. For a task, it is something you can check this month.

• Does it have an answer you can check? A model that is right nine times in ten is useful where a person checks the output, and expensive where nobody does.
• Would it take you hours? The model gets less reliable as the task gets longer, so the hours-long job is the one to leave for later.
• Does it depend on something you cannot write down? A relationship, a promise, a judgement about a person. Those do not survive being turned into a prompt.

Four gaps, stated because the numbers above get quoted constantly without them.

• Whether any of this changes pay or headcount. Every study here counts output per person, not jobs. The strongest job-level evidence comes from robots, not language models: one more robot per thousand workers went with 0.2 percentage points off the employment ratio and 0.42% off wages (Acemoglu and Restrepo, 2020).
• Whether training protects anyone's job. The World Economic Forum counts the need — 59 of every 100 workers need training by 2030, and it expects 11 of them not to get it — not the outcome.
• How often a confidently wrong answer reaches real work. The 19% figure above is one lab task, not a working week.
• Anything outside English, office work and the United States. The usage datasets do not cover it.

We would rather print those gaps than round them away. On this page, a number without a source is a number we have not earned.

Write down a week of your work as tasks, not as a job title. Then sort it into two columns.

• Moves: repeats often, has an answer you can check, follows rules somebody could write down.
• Stays: needs you in the room, needs a relationship, or needs a judgement you could not write down even if you tried.

Both columns are jobs. Taking the first route means building and supervising the automation, which looks more like consulting than operations: you bring the knowledge of how the work is really done, and you are the one who notices when the machine is confidently wrong. Taking the second means doing the human part more often and better — a position worth having only while people can find you.

That is what our people directory is for. A profile there is yours, it is public, and a person or an agent can read it; the file agents read first, llms.txt, names the directory. LinkedIn is the opposite by design. Its robots.txt tells every crawler it has not listed to stay out, then names the AI ones — ClaudeBot, GPTBot, ChatGPT-User, PerplexityBot — and refuses them the profile pages by path. A person whose only professional presence is LinkedIn is invisible to an agent that reads only what it is allowed to read. Check the file yourself.