On October 16 the Anthropic Institute published “The AI Expectations Gap,” a technical report on two surveys run in August 2026. One went to a nationally representative sample of 10,980 US adults, the other to Anthropic’s own staff. The authors are Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek and Peter McCrory.
The finding in one sentence: both groups broadly agree on what AI can do today, and split on how fast it gets better. Asked whether AI will be able to make a Nobel-level discovery by 2030, 39 percent of US adults say yes. Among staff it is 91 percent. Depending on the task, between 19 and 40 percent of adults say AI will never be able to do it. Among staff, at most 2 percent say that.
The gap widens on your own job
On average, staff expect AI to be able to do 75 percent of their current work by 2030. US workers say 28 percent. More than half of staff expect at least 80 percent; among US workers that is 5 percent. Time savings look the same: of the adults who use AI at work, 26 percent say a suitable task takes a quarter of the time or less. Among staff, 59 percent say so.
On adoption the groups sit closer together. Of the tasks AI could do by 2030, adults expect 42 percent to actually be done with AI; staff expect 65 percent. And the obstacles adults name are not technical ones. First comes customers preferring to deal with a person, then data privacy and security, then reliability, then laws and liability.
How much of model building the models do
The more interesting number is the one about themselves. Asked what share of the work of building an early-2026 frontier model AI does today, the median staff answer is 51 percent. For the end of 2027 it is 80 percent, for the end of 2030 it is 95. People on research teams put today’s share higher: 60 percent against 53 for everyone else. Anthropic is careful to say these are personal estimates from an internal poll, not a measurement and not a company forecast. It fits with Jack Clark pulling full recursive self-improvement forward from 2028 to 2027, and the numbers on employment inside the company landed the same day.
People who use AI daily think like the lab
The steadiest relationship in the study does not run between the groups but through them. People who expect AI to do more of their own work also expect higher unemployment, and at the same expected level of automation, staff and the public name about the same rate. Among adults who use AI regularly at work, 56 percent expect AI to handle at least seven of the eight listed tasks by 2030. Across all adults it is 41 percent; among people who have never touched an AI tool, 28.
So the gap is less about optimism than about exposure. Run through the model from “Economic Scenarios for Transformative AI,” even the median US adult’s answers imply the economy growing at more than twice its usual rate. Staff expectations imply GDP doubling roughly every four and a half years from 2030. Anthropic draws the gentle conclusion: public expectations will probably move as AI use spreads. The less gentle reading is that the people setting the pace are also the least representative sample of how everyone else works.