Anthropic Staff Expect Faster AI Progress Than Most Americans

Anthropic has found a striking gap between what its employees expect from artificial intelligence and what most Americans expect. In an October 16 research report, 91% of surveyed Anthropic staff said AI would be capable of a Nobel-level scientific discovery by 2030. Among surveyed US adults, 39% gave that answer. The question asked about capability, not whether an AI system would actually win a Nobel Prize.

That 52-point difference is easy to turn into a dramatic prediction. The more useful reading is that people building frontier models and the broader public inhabit different AI worlds. Both groups broadly agreed on what AI can do today. Their disagreement grows when asked how quickly it will improve, how widely it will be used and how much work it could take on within four years.

Anthropic staff expected AI to be able to perform, on average, 75% of their current work hours by 2030. Employed US respondents put that figure at 28%. More than half of staff expected AI to handle at least 80% of their work, compared with 5% of US workers. These are estimates of technical capability, not predictions that employers will eliminate the same proportion of jobs.

The difference also appears in reported time savings on tasks suited to today’s AI, such as summarising a document or drafting an email. Among US adults who use AI at work, 26% said such a task takes a quarter of the time or less with AI, including checking and corrections. Among Anthropic staff, 59% said so. Frequent workplace AI users among the public were more bullish about future capabilities than people who had never used an AI tool, but still less bullish than Anthropic employees.

That does not make staff automatically right. They may see progress earlier than most people, but they also work in an unusually AI-intensive environment. A frontier lab is not a typical hospital, school, bank or small business. And the public has reasons to distinguish between a model demonstrating a skill and an organisation trusting it with a sensitive task.

Anthropic’s survey makes that distinction explicit. US adults expected AI to be used for 42% of tasks it could technically perform by 2030, while staff expected use for 65%. When asked what might slow adoption, the public frequently pointed to customers wanting a person, privacy and security concerns, reliability and legal responsibility. Those are not simply signs of ignorance about the technology. They are practical barriers to deploying it.

The two groups were closer in how they connected their own job exposure to unemployment. People who thought AI could do more of their work also expected higher unemployment. Yet Anthropic employees largely expected their own company to keep hiring. The apparent tension reflects a crucial difference between automating tasks and reducing headcount; new products and expanding demand can create work even as individual jobs change.

The findings also fit a broader debate about what happens when powerful AI moves beyond controlled demonstrations. Better capabilities matter, but reliability, independent testing and public confidence will determine how much of that promise reaches everyday life.

There are limits to the evidence. Morning Consult surveyed 10,980 US adults online in August 2026 and weighted responses to represent the adult population. Anthropic’s internal poll drew 1,334 employees and was unweighted. The public survey was an opt-in online panel, while staff had a direct stake in and unusually close exposure to AI. A September follow-up of 7,613 US adults found 44% expected a Nobel-level discovery by 2030, up from the August figure of 39%. None of these answers establish what AI will actually accomplish. They show how far apart expectations remain, and why the next few years could bring both surprise and friction.