Anthropic compares AI expectations: 91% of its employees and 39% of Americans expect a Nobel-level discovery by 2030

Anthropic has published a study called "The AI expectations gap" on how differently people picture the future of AI. In August 2026, the company surveyed two groups: a nationally representative sample of American adults and its own employees. 91% of employees and 39% of Americans think AI will be able to make a Nobel-level scientific discovery by 2030. More than half of employees expect AI to take over at least 80% of their current work. Among American workers, 5% think so.

The study's authors are Shimon Saher, Tess Cotter, Charles Jones, Anton Korinek and Peter McCrory. The underlying question reads like an economists' debate. Forecasts of AI's impact on the economy diverge widely. The cause may be different estimates of how fast capabilities will grow, or different estimates of how quickly those capabilities will turn into real-world adoption. The survey tries to separate these two factors.

Where the forecasts diverge

Both groups think similarly about what AI can do today. Respondents were asked to name the date by which AI would perform each of eight tasks at the level of a qualified professional. The list ran from routine emails to a Nobel-level discovery. Both groups ranked the difficulty of the tasks in almost the same order.

The differences begin with the forecast horizon. About a quarter of Americans and employees believe AI can already design, run and publish a scientific study today. Within a year, 9 in 10 employees say this will be possible. Among Americans, 38% think so. Something else stands out more: depending on the task, 19% to 40% of Americans are sure AI will never master it. Among Anthropic employees, no more than 2% say this.

Experience with AI explains the gap only in part:

• 96% of employees expect AI to perform at least seven of the eight tasks by 2030;
• among Americans who regularly use AI at work (about a fifth of the sample), the figure is 56%;
• among all American adults, it is 41%;
• among those who have never used AI, it is 28%;
• 59% of employees and 26% of American workers who use AI at work say that a task AI is suited for takes a quarter of the time or less.

Their own job and unemployment

When the question turns to their own profession, the gap widens. On average, Anthropic employees expect AI to be able to do 75% of their current work by 2030. American workers overall say 28%. People who build and use frontier models every day see further and faster. Judging by the text, they also apply this to themselves without much illusion.

The most interesting finding concerns the labor market. Those who expect AI to take over a larger share of their own work also forecast higher unemployment. This relationship holds in both groups. Moreover, at the same expected level of automation of their own work, both groups name roughly the same unemployment rate. The differing forecasts, it seems, come from different assumptions about the speed of progress. The logic "my job gets automated, so unemployment rises" is the same for everyone.

In the introduction, the authors also mention a more radical scenario: future models could build their own successors. This recursive improvement, in their words, would speed up progress even more. How fast and on what scale this would happen is unclear.

How to read these numbers

The Anthropic employee sample does not reflect the views of the industry. This is a company whose business depends on a belief in fast progress. Sample sizes and selection details are not given in the available excerpt, so the groups should be compared with that caveat in mind. Still, as a snapshot of the views of the people closest to frontier models, the data is valuable: 91% on the Nobel bar shows how far the internal horizon of expectations in the labs differs from the public one.

For businesses and policymakers, the takeaway is practical. If planning for hiring, retraining or regulation rests on the expectations of rank-and-file workers, it risks playing catch-up. If it rests on the expectations of the labs, it risks overestimating the speed. It makes more sense to build scenarios at both ends of the range and track measurable signals, such as the actual share of time saved on tasks. Right now, 26% versus 59% serves as a working benchmark for checking whose forecasts start to come true.