Anthropic Surveyed Its Own Staff and the US Public on AI. The Gap Is Enormous

A new technical report from the Anthropic Institute puts two populations side by side — a nationally representative sample of US adults and Anthropic's own employees — and asks them the same questions about what AI will be able to do, how widely it will be used, and what that means for jobs. The headline result is that the two groups broadly agree on AI's current abilities and diverge sharply on the pace of what comes next. Run through Anthropic's own economic model, the median US adult's answers imply 5.4% GDP growth in 2030; the median staff member's imply 17.1%.

Agreement on today, divergence on tomorrow

The survey, fielded by Morning Consult between August 11 and 23 with 10,980 US adults, and as an internal poll of 1,334 Anthropic staff between August 9 and 12, asked respondents when AI would be able to perform eight tasks as well as a skilled professional — from routine email drafting to making a Nobel-level scientific discovery.

According to the authors, both groups ranked those tasks in nearly the same order of difficulty, and roughly a quarter of each said AI can already design, conduct and publish scientific studies. The agreement ends there. Nine in ten staff expect AI to handle that scientific-publication task within a year, against 38% of US adults. By 2030, 91% of staff expect a Nobel-level discovery from AI, compared with 39% of US adults.

The most structurally interesting finding may be the "never" category. Depending on the task, between 19% and 40% of US adults said AI will never be able to do it. Among Anthropic staff, that share topped out at 2%. That is less a disagreement about timelines than about whether there is a ceiling at all.

Usage appears to move beliefs, though not all the way. Among the roughly one-fifth of US respondents who use AI at work regularly, 56% expect AI to handle at least seven of the eight tasks by 2030 — above the 41% figure for all US adults and the 28% for people who have never used an AI tool, but far below the 96% among staff.

The adoption bottleneck is social, not technical

Anthropic separated capability from deployment, which is the right move and one many capability forecasts skip. US adults expect that by 2030, 42% of the tasks AI could do will actually be done with AI. Staff put that figure at 65%.

Asked to name the biggest obstacles to workplace AI use in 2030 from a list of nine, US adults most often chose customers preferring to deal with a person, followed by data privacy and security concerns, AI not being reliable enough, and laws, regulations or liability. In other words, the public expects preferences and rules to bind before the technology does. Regular AI users at work ranked privacy and security first instead, at 42% versus 37% among everyone else. Staff were not asked this question — a notable omission, given that the people building the systems might have the most informed view of reliability as a constraint.

On autonomy, most staff said that by 2030 AI will either complete tasks alone or with a worker checking the output. Four in ten US adults instead expect humans to do most of the work, with or without AI assistance. That distinction maps directly onto whether AI reads as a displacement technology or a productivity multiplier.

Time savings and the "so-so technology" problem

The report leans on Acemoglu and Restrepo's framing: a technology just good enough to replace workers displaces labor without raising output much. Only large productivity gains produce real growth.

Asked how long a well-suited task takes with AI versus without, 85% of staff said half the time or less, and 59% said a quarter of the time or less. Among US adults who use AI at work, the typical answer was half the time, 26% said a quarter or less, and 29% said the task takes just as long or longer — versus 5% of staff. Anthropic links staff answers to internal data it published separately, including a tripling of lines of code added by engineers in the month after adopting coding agents. Lines of code is a famously weak productivity proxy, and the report does not claim otherwise.

Own jobs, and a shared theory of unemployment

The sharpest gap is personal. Staff said, on average, that AI will be able to do 75% of their current work hours by 2030. US workers said 28%, with only one in twenty putting it at 80% or more. Among US workers in computer, engineering and science occupations who use AI regularly, the figure was 40% — closer to the national average than to the staff view.

What the two groups share is the inference they draw. Told that unemployment is currently about 4.2%, respondents who expect AI to do more of their own work forecast higher unemployment, at roughly the same rate in both samples. US workers expecting 80% or more of their work automated forecast 12.2% unemployment in 2030. General optimism about AI capabilities, holding own-job exposure fixed, was unrelated to the forecast. So the disagreement is not about economic reasoning; it is about how much of your own job you think is in scope.

One wrinkle complicates the displacement story: 95% of staff expect Anthropic's headcount to grow over the next 12 months, with a median expectation of 66% growth. Anthropic stresses these are personal views in an internal poll, not company plans. Staff were also slightly more pessimistic than the public about re-employment, with a median estimate of 12 months for a displaced worker to find a role in a new occupation.

Read the caveats

The GDP figures come from Anthropic's own "Economic Scenarios for Transformative AI" model, which converts answers on five dimensions into growth outcomes. They are not forecasts of what will happen; they are what each respondent's stated beliefs would imply if that model is right. The staff sample is unweighted and self-selected within the company, the US panel is opt-in and online, and Anthropic has an obvious commercial interest in the proposition that AI is about to transform work. The report says a September re-field of 7,613 US adults produced similar results — 5.6% implied growth, with Nobel-level expectations rising from 39% to 44%.

Also worth flagging: staff put the median share of work on building an early-2026 frontier model already done by AI at 51%, rising to 95% by end-2030. Anthropic explicitly labels this a personal estimate, not a measurement. It is a striking number with no methodology behind it.

What to watch

• Whether the expectations gap narrows as workplace AI use spreads. Anthropic's own data suggests regular users sit between non-users and staff — but still much closer to the public.
• Whether adoption friction shows up where the public predicts it — customer preference, privacy, liability — or where practitioners worry, namely reliability.
• Whether anyone independent replicates the own-job-exposure-to-unemployment gradient outside an AI lab's survey instrument.
• Whether frontier labs keep hiring aggressively while reporting that AI does most of the work, and how long those two trends coexist.
• Whether the "share of frontier model work done by AI" claim gets an actual measurement rather than a staff poll.