91% of Anthropic Staff Expect Nobel-Level AI Capability by 2030

Agreement about today, disagreement about the slope

Start with what the survey does not show. It does not show two groups living in different realities. Anthropic staff and US adults rank the same eight tasks in nearly the same order of difficulty, from routine email to a Nobel-level discovery, and they largely agree on what AI can already do. About a quarter of each group thinks AI can design, conduct and publish scientific studies today.

The split opens when the question turns to timing. By next year, 9 in 10 staff expect that full research loop to be feasible for AI, against 38% of the public. By 2030, the Nobel-level figure is 91% against 39%. At the other end of the scale, between 19% and 40% of adults, depending on the task, said AI will never match a skilled professional. The highest staff figure for "never" on any task was 2%.

Frequency of use explains part of the distance, and Anthropic says so in the report. Among adults who use AI regularly at work, 56% expect AI to manage at least seven of the eight tasks by 2030. That is well above the 41% for all adults and the 28% for people who have never used an AI tool. It is also well short of the staff figure, 96%. Regular use moves the public toward the lab. It does not close the gap.

The own-job numbers are the harder ones to dismiss

Expectations about science are easy to treat as mood. Expectations about your own desk are closer to a plan. On average, Anthropic staff said AI will be able to do 75% of their current work hours by 2030. US workers put their own figure at 28%. More than half of staff crossed the 80% line. One in twenty US workers did.

The pattern holds inside occupations that look, from the outside, similarly exposed. Workers in computer, engineering and science jobs put AI's share of their work at 33% by 2030, and regular AI users in that group put it at 40%. Both numbers sit closer to the average US worker than to the average Anthropic employee. Whatever staff are seeing, it is not simply "technical people expect more." It is specific to working where the tools are built and used all day.

Then comes the part that complicates the replacement story. Staff who expect AI to do most of their work also expect their employer to keep hiring: 95% expect Anthropic's headcount to grow over the next 12 months, the median expectation is growth of 66%, and 27% expect it to at least double. The contrast is striking, though the horizons differ. The automation answers run to 2030; the hiring answers cover only the next year. Anthropic is careful, in the report, to label the hiring figures as personal expectations from an internal poll, not company plans. A respondent can expect near-term hiring and substantial automation several years later without contradicting themselves, and the survey does not establish whether staff expect that hiring trend to continue through 2030.

One disagreement, one shared model

The most useful finding in the report is the agreement underneath. Both groups were told unemployment is about 4.2% and asked to forecast 2030. In each group, people who expect AI to do more of their own work forecast higher unemployment, at about the same rate: US workers expecting 80% or more of their work to be doable by AI forecast 12.2% unemployment on average, above the peak during the financial crisis. Anthropic interprets that similar relationship as evidence that the two groups share some assumptions about labour-market consequences, even while disagreeing about how quickly capabilities will advance. The survey does not establish that respondents hold identical economic models — only that the link each group draws between its own exposure and unemployment points the same way.

That framing also disciplines the GDP figures attached to the report. Anthropic runs each respondent's answers through its Economic Scenarios framework individually and reports the median respondent's model-implied scenario. For the respondents who answered all five input questions — 1,093 staff and 3,259 adults — the median staff respondent's answers imply annual GDP growth of 17.1% in 2030 and the median adult respondent's imply 5.4%. These are conditional model outputs based on survey responses, not GDP forecasts and not economy-wide survey medians. They are not Anthropic forecasts, and they are not the separate $44 trillion projection Jack Clark described at the IMF this weekend. They show what follows if you take each group's expectations seriously inside one set of assumptions. Taken at face value, the staff figure is the report's own illustration of an economy doubling roughly every four and a half years — that is what sustained growth at the modelled 2030 rate would mean arithmetically, not a forecast that GDP will actually double on that schedule. It is a statement about the expectations, not about the economy.

None of this settles whether staff are early or simply invested. The employee sample is valuable for understanding expectations inside a frontier lab, but it is unweighted, covers one company, and cannot represent AI researchers or technology workers generally; respondents' professional and financial interests may also shape their expectations, and the report does not pretend otherwise. The public sample is better suited to estimating expectations among US adults: 10,980 adults, weighted to the national population, with a clear finding that direct experience with the tools moves expectations a long way. If AI use at work keeps spreading, the public numbers in this report are the ones most likely to move. Watch those, not the 91%.

Sources

[1] Anthropic Institute (Sacher, Cotter, Jones, Korinek, McCrory), "The AI expectations gap," anthropic.com, published October 16, 2026, updated October 17, 2026 — primary source for all figures in this article, including the appendix's survey-sample detailsRead source
[2] Daily AI Brief, "Anthropic Staff Expect AI to Do 75% of Their Work by 2030," published October 17, 2026 — secondary cross-check of the 75% / 28%, 91% / 39% and 59% / 26% figures and the August field datesRead source
[3] AI Parabellum, "The AI Expectations Gap Between Anthropic and the Public," October 17, 2026 — secondary cross-check, including the Morning Consult field windowsRead source
[4] TechBooky, "Anthropic Staff Expect Faster AI Progress Than Most Americans," October 17, 2026 — secondary cross-check, including the October 16 report dateRead source

THE ALFW VERDICT

The gap in this report is evidence about expectations, not about what AI will actually achieve by 2030. Anthropic's staff anticipate much faster capability gains than the US public, but their numbers are personal estimates from an unweighted internal survey — not company forecasts, and not measurements of anything except belief. The public sample is nationally weighted and broader, though it carries its own survey limitations. On labour markets the two groups are closer than the headline gap suggests: in both, people who expect AI to do more of their own work also expect higher unemployment, which points to some shared assumptions about consequences underneath a disagreement about speed. The question worth tracking is whether public expectations move toward the lab's as hands-on experience spreads. In this data, adults who already use AI at work sit much nearer the staff view than adults who never touch it. Watch that gradient, not the 91%.