The workers most worried about AI taking their jobs, according to Anthropic's latest survey data, are the ones who've handed over the least work to it. The people who've automated the largest share of their tasks are the most confident about their pay, their job security, their ability to find work, and even the meaning they get from what's left.
That is the opposite of what most people would predict. And it is specific — backed by actual usage data, not self-report.
What the Survey Measured
Anthropic published the June edition of its Economic Index this month. The new piece is that they surveyed roughly 9,700 Claude users and linked their responses to behavioral logs — the actual fraction of work tasks Claude is handling for each respondent, measured from usage patterns. That number is what they call the automation share.
They asked respondents to rate the expected impact of AI on six dimensions of their work: pay, job security, ability to find a new job, meaning, autonomy, and human interaction.
Across all six, workers with higher automation shares expected more positive outcomes.
The direction you might expect — "people who've automated more fear losing jobs more" — is not what the data shows. The direction is inverted. The more of your work Claude is handling, the better you feel about where things are going.
The Selection Problem Is Real But Incomplete
The most obvious explanation is selection: optimistic people are naturally more likely to adopt AI tools aggressively. The personality type that sees opportunity and acts on it is the same type that experiments with new tools, moves early, and delegates without anxiety. So the correlation might not mean delegation causes optimism. It might mean optimism causes delegation.
That is probably partly true. But the survey data complicates simple selection as the full explanation.
The six dimensions Anthropic measured are not all the same. Pay and job security are exactly the kind of domains where optimistic, high-performing people would expect good outcomes regardless of AI. They were always going to be fine.
"Meaning" and "human interaction" are different. The fear most articulate critics of AI automation raise — at least the non-economic version — is that offloading cognitive work hollows out the experience of working. You become a QA layer for outputs you didn't think through. The interesting part goes somewhere else.
If selection were the complete explanation, you'd still expect the high-delegation group to be neutral or negative on meaning. These are people who have handed a meaningful fraction of their work to an AI. If that hollows out the experience, the effect should show up even for optimists.
Instead, the people who've automated the most report feeling more meaning in their work.
What Might Actually Explain This
There is a version of this that makes sense if you've spent time seriously integrating AI into technical work.
When you automate the parts of a job that do not require judgment — the research pass, the first draft, the boilerplate, the documentation, the test cases for cases you already understand — what remains is the stuff that requires you. The architecture decisions where the tradeoffs are unclear. The user problem where the spec hasn't been fully thought through. The debugging session where the error is in a layer nobody has documented.
If that's what's left, "meaning" going up is not surprising. You've filtered out the tasks you found tedious and kept the ones you found interesting. That is not an argument for reckless delegation. But it might explain why the people doing the most delegation feel better about work, not worse.
The autonomy finding tracks with this too. When AI handles the scaffolding, you spend more time deciding what to build and less time executing on decisions someone else could make. More time on direction, less on implementation. That tends to feel like autonomy even when you're directing AI rather than doing the work yourself.
The Weekend Pattern
One of the other things Anthropic tracked in this report is cadences — how AI usage changes across the week.
The Claude Code pattern on weekends is telling. Weekday usage skews toward what you'd expect from a developer at a job: backend architecture, API debugging, code review. Weekend Claude Code usage shifts toward AI agent design, quant trading systems, gaming, and — this one is specific — starting a business. Job applications fall sharply on weekends. Starting a business peaks.
The people using AI the most aren't just more optimistic about their jobs. They're building things on the side. The tool didn't just make them confident in their current employment — it appears to have made them entrepreneurial in their off-hours. They're using the same capability that handles their day job to prototype the next thing.
That is not what you'd expect from people anxious about being displaced. It's what you'd expect from people who see the tool as an opening.
The Data Has a Recency Problem
One honest caveat: the Anthropic survey asks about expected outcomes over the next twelve months. More than a third of respondents expected significant changes to their responsibilities, while about 10% expected job loss as likely.
Twelve months is a short window. The previous Anthropic research on skill development from AI delegation — the February study on junior developers learning an unfamiliar library — found that comprehension gaps from heavy delegation don't announce themselves in the first few months. They show up when a debugging session requires first-principles reasoning about code you never thought through.
So the June survey and the February study are measuring different things on different timescales. Both can be accurate at once. Right now, in mid-2026, heavy AI delegators feel good about their careers and have real reasons to. Whether that optimism holds for developers who've built shallow mental models on a foundation of delegated outputs is a different question with a later answer.
The optimism in the survey is real. It's just measuring a short window in a story that's still running.
What This Changes
The practical reading of the Anthropic data is not "delegate everything." It is something closer to: the fear of AI isn't correlated with how much AI you've been exposed to. It's correlated with how little you've engaged with it.
That has implications for how organizations should think about internal AI adoption. The anxiety in most enterprise settings comes from employees who've been told AI is coming rather than given tools and time to run it themselves. Exposure does not increase fear — it seems to reduce it. The optimism gap between high and low delegators in the Anthropic data runs across all six measured dimensions in a way that suggests the relationship is with the tool itself, not with general optimism as a personality trait.
At xeve, we track which tools actually occupy your hours — not which ones you say you use. Looking at your coding tool time versus your AI assistant time versus your browser time over weeks gives you the actual ratio, not the self-reported one. The developers who feel most confident about where they're headed in the Anthropic data are the ones for whom that ratio is already shifted.
The survey asked what they expected. It did not ask why. But the simplest reading of the pattern is that if you want to be less anxious about AI, using it more is the better path — not as a consolation, but as calibration. The people who know the tool best, because they've handed it the most, are the ones least surprised by what it can and can't do. That knowledge appears to be where the confidence comes from.