The people closest to the machine are worried about the people below them

Anthropic surveyed 9,700 of its own active users and found that the workers delegating the most to Claude are the most optimistic about their own pay and the most worried about junior colleagues. That gap is the story. It is not a story about how workers feel about AI in general. It is a story about a workforce that is quietly sorting itself into people who use the machine and people the machine is used on.

The report is called "Cadences." It was published on 26 June 2026 and is the sixth Anthropic Economic Index. ¹ For the first time, the company linked hourly usage logs from Claude.ai, Claude Code, and its agent product Cowork to a survey of active users, so it could compare what people do with what people say. The headline finding, in Anthropic's own framing, is that the heaviest delegators feel best. They expect better pay. They feel more secure. They report more meaning in their work.

Read the second number, not the first. The same group of heavy delegators is also the most worried about what AI will do to the jobs of more junior people in their field. ^{1 2} Lower-delegation users do not show that split. The people closest to the machine are the people most willing to say, on the record, that the rung below them is being pulled up.

What the sample can and cannot tell us

Before the argument, the obvious caveat. This is a survey of active Claude users. It is not a survey of the workforce. It is not a survey of people who lost a job to AI last quarter, or of people whose employer has not bought them a seat, or of people whose work was never on Anthropic's roadmap in the first place. ExplainX, summarising the methodology, makes the non-representativeness explicit. ² The people in the sample have already chosen to integrate Claude into their working week. They are pre-selected for a positive relationship with it.

That matters for how we read the optimism number. It does not matter as much for how we read the worry number. When a self-selected, AI-fluent, professionally confident sample tells its own surveyor that it is worried about the juniors, the worry is doing work against the grain of the sample's bias. It is the finding that survives the sampling problem.

The reframing

The dominant frame around AI and work is still individual. Will it take my job. Will it raise my wages. Will I be augmented or replaced. Anthropic's report, and most of the commentary around it, sits inside that frame. Heavy delegators are bullish about themselves, so the augmentation story holds.

The structural question Cadences accidentally surfaces is not whether senior knowledge workers are being augmented. It is what happens to the pipeline that produced them. Junior roles in law, software, consulting, research, and design have historically done two things at once. They have produced output, and they have produced senior workers. The output part is the part Claude is good at. EdTech Innovation Hub, summarising the same dataset, reports the speed-up is sharpest at college-level cognitive tasks, on the order of a twelvefold acceleration. ³ That is exactly the layer where juniors used to learn their craft by doing the work slowly and badly until they did it quickly and well.

If senior workers absorb that work through delegation to a model, the output gets produced. The senior workers get faster. The juniors do not get hired, or get hired in smaller numbers, or get hired into roles that no longer contain the developmental tasks. The senior workers in Anthropic's sample appear to see this. They are telling the surveyor about it.

Who pays

The distributional reading is not subtle. Anthropic's own numbers say the productivity dividend concentrates at college-level tasks and enterprise API use. ³ That is the upper part of the wage distribution. Workers doing routine low-skill work do not get a twelvefold speed-up; in many cases their jobs were not on the roadmap to begin with, because the economics of building a model to do them did not pencil. The gains flow up.

Inside the professional tier, the same logic repeats one level down. Senior workers with the judgement to delegate well capture the speed-up. Junior workers, who do not yet have that judgement, are the ones whose tasks the seniors are absorbing. The distributional picture, on Anthropic's data, is not workers versus capital. It is a layered enclosure: each tier of skilled labour is capturing some of the dividend from the tier below.

Daron Acemoglu's macroeconomic work on AI argues that historical automation episodes have, on net, transferred gains to capital and high-skill labour rather than lifting median wages. ⁴ Cadences does not engage that literature. It does not need to. The pattern it documents at the level of individual user sessions is the same pattern, observed at a finer grain.

On who is doing the measuring

I do not want to dismiss the data because Anthropic produced it. The telemetry is real. The 9,700 users are real. The survey instrument, by the standards of corporate research, is not obviously rigged. But Anthropic designed it, fielded it, defined what counts as "augmentation," chose which findings to lead with, and published it under a frame that emphasises optimism. The cigarette-company-on-smoking-research comparison is too strong; the pharmaceutical-trial-run-by-the-manufacturer comparison is about right. You can use the numbers. You should not let the company narrate them for you.

The narration choice in Cadences is to lead with the optimism of heavy users and to treat the junior-worker worry as a curiosity. Read the other way round, the report is Anthropic's own data telling Anthropic's own customers that the AI-fluent professional class expects the rung below them to be cut. That is not a marketing finding. It is a labour-market finding the company has politely placed in the second half of the paragraph.

The thing to watch over the next year is not whether the optimistic users were right about their own pay. They probably were. The thing to watch is hiring at the junior end of the professions where delegation is highest. If those roles shrink, or get redesigned into something that no longer trains anyone, the senior workers in Anthropic's sample will have been the first reliable narrators of a pipeline that is closing behind them.

Footnotes

¶