Three findings from the same survey, in the order they were reported.

First: 57% of respondents said AI had made their skills more valuable, and 68% said they were learning more with it (Anthropic Economic Index, June 2026).

Second: “the share of people reporting that AI is increasing the market value of their skills rises with automation share, while the share reporting they learn more is roughly flat” (Anthropic Economic Index, June 2026).

Third: those two sentences describe curves with different shapes, and almost every write-up of this report has collapsed them into one.

The survey linked responses to actual usage data for roughly 9,700 people between mid-May and early June 2026 (Anthropic Economic Index, June 2026). “Automation share” is the fraction of a person’s conversations where the work was handed over rather than worked through together. So this is not people estimating how much they delegate. It is what they actually did, set against what they believe about themselves.

The two beliefs come apart

Hand over more, and you are more likely to say your skills are worth more. Hand over more, and you are no more likely to say you are learning.

Both of those can be true at once, and there is a perfectly reasonable reading in which nothing is wrong. Skills become more valuable when they are multiplied. A competent editor with a fast drafting tool ships more than a competent editor without one, and the market pays for shipped work rather than for effort. On that reading, “my skills are worth more” is a claim about reach, not about growth, and the flat learning line is simply irrelevant to it.

There is a second reading that is harder to dismiss. Delegating a task and understanding a task are separable, and the survey shows people separating them. If your sense that your skills are appreciating is coming from output volume, and your understanding of the work is not moving, you are describing a position that depends on the tool remaining available, remaining cheap, and remaining better at the task than the next person’s tool. That is a real position. It is just not the same thing as your skills being worth more.

The data cannot tell you which reading is right. It is a cross-section, not a controlled trial: nobody was assigned a delegation level, and the people who delegate most may simply be the people whose work suits it. Causation is unavailable here in every direction.

What the data does establish is narrower and still useful. The two beliefs are not interchangeable, and the survey population is treating them as separate things. If you have been justifying how much you hand over with “I’m learning from watching it work,” this is the finding that says check that, because at the population level the learning line does not rise with the handing over.

The context around it is less cheerful

Two other numbers from the same report are worth putting beside the optimism, because they are measuring different people.

Only 10% of respondents “rated losing their own jobs as likely or very likely” in the next twelve months. Concern for junior colleagues ran much higher (Anthropic Economic Index, June 2026). And workers in their first year of work reported task shares that AI can already perform roughly 10 percentage points higher than workers with fifteen or more years of experience (Anthropic Economic Index, June 2026).

Stack that against the first finding and the shape is clear enough. The people reporting that AI makes their skills more valuable are, on this evidence, disproportionately the people who already had skills to multiply. The exposure is concentrated at the other end of the tenure curve, among people who have not yet built the thing the tool is supposed to amplify.

That is not a prediction about anyone’s job. It is a statement about where the two effects are landing, and they are not landing on the same people.

One thing to do this week

Pick a single task you now hand over by default: the weekly summary, the first draft, the data pull. Do one instance of it yourself, end to end, and time it.

You are not testing whether you are faster. You know you are not. You are testing whether you can still do it well, and whether your judgement about the output has outrun your ability to produce it. If the answer is comfortable, your delegation is doing what you think it is. If it is uncomfortable, you have found the specific thing to keep your hands on, and it costs one afternoon a quarter to keep it.

A note on the source. This is Anthropic’s own research, drawn from usage of its own product, and self-reported beliefs at that. That places it below a government statistical release on any reasonable scale of independence, and we have treated it accordingly. Every figure above is quoted as the report states it, with the sample and the measure named, rather than presented as a settled fact about work in general. The underlying data is published openly, which is more than most claims about AI at work can say. This site’s approach to sourcing and to its own use of AI tools is set out in our editorial policy.