One in Five Workers Says AI Has Already Taken Parts of Their Job. Here Is What They’re Not Saying.

The debate about whether AI is actually affecting jobs has, for the past three years, been conducted mostly in the abstract. Projections. Models. Theoretical exposure scores. Economists arguing about whether the data was showing anything yet.

A survey published this week by Epoch AI and Ipsos has put something more direct on the table. Of 2,000 American adults surveyed, 20% of full-time workers said that AI has already replaced tasks they previously did as part of their job. Not might replace. Has replaced. Already. Past tense.

15% said AI had created new tasks they would not have done otherwise.

The gap between those two numbers — five percentage points — is the story that is not being told.

The survey was conducted by Ipsos in partnership with Epoch AI, a nonprofit focused on data-driven research about AI development and impact. It found that half of American adults used AI in the previous week for either personal or work purposes. ChatGPT was the most widely used service, reported by 31% of respondents, followed by Google’s Gemini at 21% and Microsoft’s Copilot at roughly 10%.

These are not fringe numbers. Half of American adults. One in five full-time workers with tasks already displaced. These are mainstream figures describing a mainstream shift that is underway right now, not in 2030.

What the survey captures, and what most coverage of it has understated, is the asymmetry.

When economists model AI’s impact on employment, they typically note that technology both destroys and creates jobs — and that historically, the creation has outpaced the destruction over long time horizons. That is true. It is also, for the purposes of understanding what is happening to people right now, not particularly useful.

The 20% who say AI has taken parts of their job are experiencing that displacement today. The new tasks and new roles that AI creates will materialise over months and years, require different skills, and in many cases go to different people. The person whose document review work was absorbed by an AI system this quarter is not automatically positioned for the AI oversight role that might emerge in six months. The destruction and the creation are not happening to the same people on the same timeline.

Goldman Sachs put the net figure at 16,000 jobs lost per month when substitution is weighed against augmentation across the US economy. The Epoch AI survey suggests the subjective experience of displacement is considerably broader than any monthly net figure captures. Twenty percent of the full-time workforce is not 16,000 people. It is tens of millions.

Half of American adults used AI in the past week. Most of them, presumably, did not think of that usage as participating in their own displacement. They were using a tool. Answering an email faster. Summarising a document. Generating a first draft. These feel like efficiency gains. And they are. They are also, in aggregate, the mechanism by which the 20% figure is produced — by which tasks that once required sustained human attention are absorbed into workflows where they no longer do.

There is an honest question embedded in the 15% figure — the workers who say AI has created new tasks for them — that deserves more attention than it typically receives.

What are those new tasks? In many cases, they are tasks that involve supervising, correcting, or directing AI output. Reviewing what the model generated. Catching what it got wrong. Redirecting it when it drifts. These are real tasks that require human judgment, and they are not trivial. They are also, as we explored in the piece about older workers retraining as AI data annotators, often paid significantly less than the work they replaced.

The new tasks AI is creating are real. They are not equivalent, in skill or compensation, to the tasks it is replacing. That is the part of the story that the 15% figure does not tell by itself.

What this survey adds to the picture that was already forming from Goldman Sachs, from Anthropic’s research, from the Q1 tech layoff data, is a human dimension. Not model outputs or regression coefficients — actual people, in actual jobs, reporting what has happened to their working lives.

One in five. Already. In a survey conducted this week.

The debate about whether AI is affecting jobs is over. The debate that matters now is about what happens to the people it has already affected, and what happens next to the much larger number it has not yet reached but will.