Anthropic’s latest study finds a striking difference between what its employees and the wider US public expect artificial intelligence to accomplish by 2030. The gap is clearest on frontier capability and work: 91% of Anthropic staff say AI could make a Nobel-level discovery by then, compared with 39% of US adults, while staff expect AI to perform 75% of their current work hours versus 28% among employed adults. Those are expectations, not demonstrations of what a system can do or forecasts of what will necessarily happen.
What Anthropic measured
Anthropic’s technical report, published October 16, 2026, compares two distinct groups rather than treating them as interchangeable. The public portion surveyed 10,980 US adults through an opt-in Morning Consult online panel. Those responses were weighted, but an opt-in panel is not the same thing as a probability sample. The second group was an internal, unweighted poll of 1,334 Anthropic staff members. The sampling and weighting distinction matters because the public results are intended to describe reported expectations in a weighted survey sample, while the staff results describe the views of people inside a frontier AI company. Anthropic’s primary report and its statistical appendix provide the study framing and sample details.
The questions addressed beliefs about future capabilities, the amount of respondents’ work that AI might perform, unemployment, and economic growth. The central comparison is therefore a comparison of expectations. It is not a test in which an AI system attempted a Nobel-level discovery, completed a representative set of jobs, or generated the GDP outcomes described later. Keeping that distinction in view prevents the headline numbers from becoming stronger claims than the study supports. Anthropic’s report describes the results as survey evidence about how the two groups see the future.
The capability divide
The most memorable result concerns scientific discovery. Ninety-one percent of Anthropic staff said AI could make a Nobel-level discovery by 2030, compared with 39% of US adults. The figures show a 52-percentage-point difference in belief about the pace of progress. They do not establish that an AI will make such a discovery, identify what “make” would mean in practice, or measure whether current systems possess the required ability. They record what respondents expect by a specified date. The primary report presents the comparison as an expectations gap.
The contrast is notable because the two groups are not simply two random slices of the same population. Anthropic staff work at an AI company and participated in an internal poll; the public respondents came from a weighted, opt-in US adult panel. That context can shape what people have seen, used, or inferred about the technology, but the supplied study results do not identify a single cause for the difference. The most defensible conclusion is narrower: staff and public respondents report materially different beliefs about the speed and scale of AI progress. The appendix should be consulted for the survey instruments and estimation details.
Business Insider used a more dramatic headline, saying Anthropic employees were more sure than most people that AI would “take over jobs.” That is useful as a record of secondary coverage, but it is not the wording to substitute for the primary study. The primary result is about what respondents think AI may be able to do by 2030, not a claim that AI will take over all jobs or that job loss is inevitable. Business Insider’s report supplies the secondary framing; Anthropic’s report supplies the underlying survey language.
The work-hours gap
Respondents were also asked how much of their own current work AI could perform by 2030. Anthropic staff gave an average answer of 75% of work hours. Among respondents who reported having a job, the corresponding US adult average was 28%. Only about 5% of US workers expected AI to perform at least 80% of their work hours. The comparison is large, but the question asks about perceived future task performance, not whether employers will automate that share, eliminate the associated roles, or preserve the same jobs with new tools. Anthropic’s findings report all three figures.
The denominator also deserves attention. The 28% figure applies to employed respondents in the public sample, while the 75% figure describes Anthropic staff. The public and staff samples should not be read as if they were two equally representative measures of all workers. The internal staff poll was unweighted, and the public survey came from an opt-in online panel, even though the public responses were weighted. These design choices do not erase the reported difference, but they limit how broadly it should be generalized. Anthropic’s appendix is the relevant source for interpreting those samples.
The report separates capability expectations from the broader question of adoption. A respondent may believe AI can perform a task without believing a firm will deploy it, that a customer will accept it, or that a worker will be replaced. Conversely, a tool can change a job without performing every hour associated with that job. Those are interpretive distinctions rather than additional survey results, but they explain why “AI can perform 75% of hours” should not be rewritten as “75% of jobs will disappear.” The primary report does not make that equivalence. Read the primary wording here.
Unemployment: an association, not a forecast
The survey connected expectations about personal work exposure with respondents’ forecasts for the US unemployment rate in 2030. Among US workers who expected AI to perform at least 80% of their work, the mean unemployment forecast was 12.2%. That is a mean response within a subgroup, not Anthropic’s forecast of the unemployment rate and not evidence that high expected exposure causes high unemployment. Anthropic’s report explicitly presents this relationship as an association within the survey.
This distinction is especially important because the 12.2% figure can sound like a macroeconomic prediction when detached from its denominator. It describes the average 2030 unemployment rate forecast among workers who selected the highest exposure category—at least 80% of their own work. It does not say that the United States will reach 12.2%, that 80% of work will be automated, or that one outcome mechanically produces the other. A respondent’s beliefs about capabilities, labor markets, business decisions, and future policy may all be entangled; a cross-sectional survey cannot by itself separate those pathways. The appendix is the source to use for the grouping and analysis details.
The pattern appears in both groups: people who expect AI to do more of their own work also report higher unemployment expectations. The report’s point is that staff and public respondents appear to have different beliefs about how quickly AI capabilities will advance, while their reported connection between personal exposure and broader unemployment is similar. That is an interpretation of the survey pattern, not proof of a labor-market mechanism. The primary analysis supports the narrower description.
The conditional GDP scenarios
Anthropic also mapped answers to a structured economic model. The model-implied median GDP growth scenario for the public sample was 5.4% in 2030, compared with 17.1% for the staff sample. These figures are conditional outputs: they represent the median scenario implied by respondents’ answers to the model’s inputs, rather than a company forecast, a consensus forecast, or an observed growth rate. Anthropic’s report reports the two medians and their conditional status.
The model translates survey beliefs across dimensions such as AI capabilities, diffusion, automation, productivity, and worker transition frictions. It does not settle whether those beliefs will prove accurate. More importantly, the scenario is not a complete account of the economy: the stated caveats include policy, business cycles, demand conditions, and market shocks. Because those forces can change outcomes independently of respondents’ technology expectations, the 5.4% and 17.1% values should be read as model-implied scenarios under specified assumptions. The appendix and primary report provide the appropriate caveat.
What the study can—and cannot—tell us
The study’s strongest finding is descriptive: Anthropic staff are much more optimistic about the speed of AI progress and much more likely to expect AI to perform a large share of their own work by 2030 than the surveyed US public. It also shows that workers with higher expected personal exposure report higher unemployment expectations. Those findings can inform debate about how differently exposed groups view the transition, but they are not measurements of future capability, productivity, employment, or GDP.
Readers should keep four limits beside every headline number. First, the public sample came from an opt-in panel, even though responses were weighted. Second, the staff poll was internal and unweighted, so it is not a representative survey of workers or even necessarily of every Anthropic employee. Third, survey expectations are not outcomes. Fourth, the GDP exercise is a conditional model that omits or does not fully represent policy, business cycles, demand, and market shocks. These caveats do not make the results meaningless; they define what the evidence can support. The primary report and appendix should remain the reference points.
A reasonable reading is that the technology’s closest observers anticipate a faster transition than many members of the public do. Whether that expectation becomes reality will depend on capabilities, deployment, institutions, demand, and labor-market adjustment—questions this survey identifies but does not resolve. That is why the useful news is not that one side has proved the future. It is that the two sides are starting from very different expectations about 2030. Anthropic’s study documents the gap; it does not close it.
FAQs
Did Anthropic prove that AI will make a Nobel-level discovery by 2030?
No. The 91% figure is the share of Anthropic staff who said AI could make a Nobel-level discovery by 2030; the comparable US adult figure was 39%. It measures belief, not demonstrated ability or a verified prediction. See the primary report.
Does the study predict that 75% of jobs will disappear?
No. Anthropic staff expected AI to perform 75% of their current work hours by 2030, while employed US adults expected 28%. The question concerns perceived ability to perform work, not job losses, employer decisions, or the future number of jobs. Anthropic’s work-hours findings make that distinction.
Is 12.2% the report’s unemployment forecast?
No. It is the mean 2030 unemployment forecast among US workers who expected AI to perform at least 80% of their own work. The result is an association within the survey, not a causal estimate or an official forecast. Read the unemployment analysis.
What do the 5.4% and 17.1% GDP figures mean?
They are model-implied median 2030 GDP growth scenarios based on answers from the public and staff samples, respectively. They are conditional outputs, not Anthropic’s company forecast, and the model does not capture every policy, demand, cycle, or market shock. The primary report explains the exercise.
Can the staff and public samples be compared directly?
They can be compared descriptively, with care. The public sample contains 10,980 US adults from an opt-in, weighted Morning Consult online panel; the internal staff poll contains 1,334 unweighted Anthropic employees. Their different designs mean the contrast should not be generalized as though both were representative probability samples. Consult the appendix.