91% of AI Employees Expect Nobel-Level Discovery by 2030—Only 39% of the Public Does

91% versus 39%. In an Anthropic survey published on October 16, 2026, nine in ten employees said AI could be capable of making a Nobel-level scientific discovery by 2030. Among a weighted sample of U.S. adults, 39% expected the same.

Those numbers are not the probability that an AI will win a Nobel Prize, nor are they Anthropic’s official business forecast. The question asked when respondents expected AI to possess the capability to perform a task—not when companies would actually deploy it or institutions would accept the result. The survey measures an expectations gap, not a prophecy.

Who was asked, and what were they asked?

Anthropic surveyed U.S. adults and its own employees in August 2026. The public sample included 10,980 adults and was weighted to represent the national population. Employee responses were not weighted. The survey described eight levels of capability, ranging from writing everyday email to designing, executing, and publishing scientific research and making a Nobel-level discovery.

The two groups broadly agreed on the ordering of what AI can do today. They diverged on the speed of future progress. Roughly a quarter of each group thought AI could currently design and publish scientific research, but 90% of employees expected that ability by 2027, compared with 38% of the public. On Nobel-level discovery by 2030, the gap was 52 percentage points: 91% versus 39%.

Americans who regularly used AI at work were more optimistic than nonusers. The share expecting AI to perform at least seven of the eight tasks by 2030 was 56% among regular workplace users, 41% among all adults, and 28% among people who did not use AI tools. Even the regular users remained below the 96% employee result on that combined threshold.

That middle tier is worth pausing on, because it complicates the simplest reading of the results. If the 52-point gap were purely a story of insiders versus outsiders, regular workplace users should sit closer to the employee end of the spectrum. They do not. They are measurably more optimistic than nonusers and measurably less optimistic than the people who build the models, which suggests that exposure to AI in daily work shifts expectations without closing the distance between users and makers. The ladder from nonuser to user to employee is real, but each rung is far apart, and the top rung is occupied by a group with unusually strong reasons to look upward.

Employees are not an independent expert panel

The comparison group is the survey’s biggest caveat. Anthropic employees build frontier models and use them every day. They may see technical progress earlier and in more detail than the public. At the same time, the company’s mission, hiring, incentives, and product culture can select for and reinforce optimistic expectations.

The researchers do not present the employee survey as an official Anthropic plan. Employees are not a randomly sampled panel of AI experts, and they do not come from the same population as the nationally weighted public sample. Reading the 91–39 split as “the experts are right and the public is wrong” goes beyond what the study design can support.

“Nobel-level” also describes the scientific significance of a result. It does not answer whether the Nobel committee would recognize an AI, whether humans would be coauthors, or how verification and responsibility would work. A real discovery must pass bottlenecks involving physical laboratories, data, replication, peer review, regulation, and funding.

Those bottlenecks are not incidental details to be waved away once a model clears a benchmark. A capability threshold describes what a system can do under favorable conditions; a published discovery describes what survived contact with the world. Experiments must be run, often by people and machines that the model does not control. Results must replicate. Reviewers must be persuaded. Institutions must decide who is accountable when a finding turns out to be wrong. Each of these steps has its own timeline, and none of them appears in a survey question about capability. The 91% figure can be entirely correct about what AI will be able to do while the pace of actual scientific output remains governed by processes the model never touches.

The productivity and employment forecasts diverged too

The gap was not limited to science. When working with AI on tasks where it performs well, 59% of employees said completion time falls to one quarter or less. Among U.S. adults who use AI at work, 26% reported the same. Conversely, 5% of employees versus 29% of workplace users said the task took the same amount of time or longer.

The difference was even sharper when respondents considered their own jobs. On average, employees expected AI to be capable of performing work equivalent to 75% of their current hours by 2030. The average among U.S. workers was 28%, and only 5% of workers thought the figure would reach at least 80%.

But “AI can perform a share of tasks” is not the same as “that share of employees will be fired.” Ninety-five percent of employees expected Anthropic’s headcount to grow over the following 12 months. That was an individual expectation, not a disclosed hiring plan. Automation can reduce labor per task while higher output and demand increase total employment.

The mechanism behind that apparent contradiction is familiar from earlier waves of automation. If a tool lowers the cost of producing a unit of work, an organization can produce more units, serve more customers, or pursue projects it previously could not afford. Headcount can rise even as the labor required per task falls, provided demand expands fast enough. The employees’ optimism about their own employer’s growth is therefore not inconsistent with their belief that AI will absorb most of their current hours. It is a bet that the hours freed up will be redirected rather than eliminated—a bet that the survey records but cannot verify.

The 17.1% GDP figure is a scenario, not a forecast

The researchers fed survey answers into a separate economic model. Using median responses, the public’s expectations produced a scenario with 5.4% U.S. GDP growth by 2030, while employee expectations produced 17.1%.

These are not observed growth rates or Anthropic’s macroeconomic predictions. They are conditional outputs produced by combining assumptions about capability, adoption, automation, time savings, and reemployment friction.

The model also did not use every survey respondent. Its core analysis included 3,259 U.S. adults and 1,093 employees who completed the necessary questions. Different model structures or different expectations fed into the same structure could materially change the result.

The word “conditional” is doing a lot of work here. A scenario of 17.1% growth is not a claim that the economy will grow that fast; it is a statement about what would follow if the employees’ expectations about capability, adoption, and time savings were fed through a particular set of economic relationships. Change the assumptions about how quickly firms reorganize around new tools, or about how much friction workers encounter moving between roles, and the same survey answers can yield a very different number. The gap between 5.4% and 17.1% is best read as a measure of how much the two groups’ expectations diverge once they are translated into a common framework—not as a range within which the future is likely to land.

What the gap actually tells us

First, frequent AI users expect faster progress. The survey cannot determine whether that reflects better information, selection bias, or both.

Second, technical capability and social adoption are separate. U.S. adults identified customers’ preference for humans, privacy and security, reliability, law, and accountability as major barriers. A model can become capable before an organization is willing or permitted to use it.

Third, both groups associated greater automation of their own work with a higher 2030 unemployment rate. Their capability forecasts differed, but they connected personal task change to labor-market risk in similar ways.

That third finding deserves emphasis because it cuts against the intuition that optimism and anxiety travel together. The groups disagreed sharply about what AI will be able to do, yet they agreed on how personal automation relates to labor-market risk. The divergence is about capability and timing, not about whether change will be disruptive. Both groups appear to accept that their own tasks will change; they differ on how soon and how completely.

For the reader, the practical implication is to separate three questions that the survey deliberately keeps apart. What will AI be able to do? When will organizations actually use it? And what will that mean for output, prices, and jobs? The 91–39 split answers only the first, and only as an expectation. The productivity figures address the second from the perspective of people who have already adopted the tools. The GDP scenarios attempt the third under explicit assumptions. Treating any one of these numbers as a prediction of the others is the most common way to misread the study.

Verdict

The 91% figure does not mean Anthropic employees assign a 91% probability to AI receiving a Nobel Prize in 2030. It means that share expected AI to cross a particular capability threshold by then. The 39% result does not mean the public dismisses AI. Nearly four in ten U.S. adults expected Nobel-level discovery capability within four years.

The more surprising result may be that both groups expect rapid change. The real disagreement is not whether AI will improve, but how quickly capability becomes real scientific output, productivity, and employment change. The survey cannot settle that question. It does show that the people building the technology and the people most affected by it already hold sharply different expectations.