1. What the AI expectations gap report found
On October 16, 2026, The Anthropic Institute published a technical report titled "The AI Expectations Gap," comparing how the company's own employees and the American public expect artificial intelligence to change work. The central result: Anthropic staff expect AI to be able to do 75% of their current work by 2030, while US workers overall expect 28%.
The report was written by Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek and Peter McCrory. It asks a simple question with large consequences. When forecasts of AI's economic impact differ so widely, is the disagreement mostly about how quickly the technology improves, or about how far improvements translate into real-world adoption?
The authors note that forecasts of AI's effects vary enormously and are fiercely debated, in part because systems that help build better systems could speed progress. Their approach was to put identical questions to two very different groups and then trace where the answers diverge.
2. How was the survey conducted?
Two main samples were used. The public sample came from an online opt-in panel run by Morning Consult, fielded between August 11 and 23, 2026. It included 10,980 US adults, weighted to be nationally representative, though it necessarily leaves out adults without internet access. Half of those respondents were asked about 2027 and half about 2030; the 2030 answers are the ones reported unless stated otherwise.
The second sample was an internal poll of Anthropic employees, conducted from August 9 to 12, 2026. A total of 1,334 staff from across the company responded. Their answers were not weighted, and all of them were asked about 2030.
Two further checks were added. Morning Consult repeated the public survey from September 17 to 24 with 7,613 respondents and two extra questions, and the results moved little: the median implied GDP growth figure was 5.6% rather than 5.4%, and the share expecting a Nobel-level AI discovery by 2030 rose from 39% to 44%. Separately, a scenario explorer on the Anthropic website, launched on September 9, drew 23,075 completed quizzes from more than 140 countries within 48 hours, although those visitors chose themselves and are probably more interested in AI than the average person.
3. Where do staff and the public agree, and where do they split?
Both groups broadly agree on what AI can do now and rank tasks in a similar order of difficulty. The split opens when the question turns to the future. About a quarter of each group says that AI can already design, run and publish scientific studies. Asked about the following year, 90% of Anthropic staff expect that to be possible, against 38% of the public.
The widest gaps appear on the most ambitious milestones. For a Nobel-level scientific discovery by 2030, 91% of staff say AI will get there, compared with 39% of US adults. Among the public, between 19% and 40% say that AI will never be able to perform a given task on the report's list, whereas no more than 2% of staff say so.
Experience with the technology narrows the distance. Roughly one-fifth of respondents use AI regularly at work, and 56% of them expect AI to handle at least seven of the eight listed tasks by 2030. That compares with 41% of all adults and 28% of people who never use AI. Among Anthropic staff the figure is 96%.
4. How much of their own jobs do respondents expect AI to do?
Asked what share of their current working hours AI could perform by 2030, Anthropic employees averaged 75%, and more than half said 80% or higher. US workers averaged 28%, and only about 5% placed their answer at 80% or above.
Occupation matters among American workers. Those in manual trades, production and transportation averaged 25%, while workers in computer, engineering and science jobs averaged 33%. Regular AI users within that technical group expected 40%, closer to the staff view than the rest of the public, though still well below it.
The authors also distinguish between what AI can do and what will be done with AI. By 2030, the public expects 42% of AI-capable tasks to actually be performed with AI, while staff expect 65%. On the question of how work will be organized, most Anthropic employees foresee AI finishing tasks alone or with a worker checking the result, whereas about four in ten US adults expect humans to do most of the work themselves or with some AI help.
5. What slows adoption, and does AI save time at work?
Members of the public were asked to choose up to three obstacles from a list of nine. The most frequently selected were customers preferring to deal with a person, concerns about data privacy and security, AI not being reliable enough, and laws, regulations or liability. People who use AI at work ranked privacy first, at 42% compared with 37% of other respondents, and gave less weight to customer preference, at 37% against 42%.
Anthropic staff were not asked this question, so the report cannot say how the company's employees would rank the same hurdles. That leaves one half of the adoption story as the public's view of friction, with no direct staff counterpart.
Time savings produced another stark contrast. Among staff, 85% said that tasks suited to AI take half the time or less once checking and correcting are included. The typical American who uses AI at work also answered half the time, but 29% of them reported no savings or even longer completion times, compared with 5% of staff.
At the more extreme end, the share saying AI-suited tasks take a quarter of the time or less was 26% among US adults who use AI at work and 21% among those who do not, versus 59% of Anthropic employees. The report points to internal observations to explain the staff view, including a tripling of engineers' lines of code in the month after they adopted coding agents and more deals closed by sales representatives using agents. Those observations come from a companion Anthropic report on employment at the company.
6. What do they expect for unemployment?
Respondents were told that the current unemployment rate is about 4.2% and asked to forecast the 2030 rate. The answers track personal exposure to AI. US workers who expect AI to do less than half of their work forecast unemployment of 7% to 8%, while those who expect it to do 80% or more forecast 12.2%, a level above the peak of the Great Financial Crisis.
Each additional 10 points of expected exposure was associated with roughly 0.4 percentage points more in forecast unemployment. Once a respondent's own job exposure is taken into account, general optimism about AI's capabilities does not predict the unemployment forecast. Staff display the same pattern, and at equal exposure levels their forecasts closely match those of the public.
That finding matters for how the gap should be read. The difference in headline unemployment forecasts between the two groups largely reflects how much of their own work each group expects to be automated, rather than a different underlying view of how labor markets respond.
Respondents were asked how long a worker displaced by AI in 2030, and forced to change occupation, would need to be re-employed. The staff median was 12 months. The US adult median fell within the range of 7 to 12 months. For comparison, the report cites a baseline of about three months in normal times.
Employees also offered a view of their own employer's future. Some 95% expect Anthropic's headcount to grow over the next 12 months, with a median expected growth of 66%, and 27% expect it to at least double. The authors stress that these are personal opinions given in an internal poll, not company plans.
7. What do the answers imply for economic growth?
To turn the survey answers into a common yardstick, the authors fed them into an economic model that covers five dimensions: capabilities, diffusion, automation, productivity and the frictions workers face in transitioning. It builds on their separate working paper, "Economic Scenarios for Transformative AI," and holds other parameters at the values of its "Substantial" scenario.
The median US adult's answers imply 2030 GDP growth of 5.4%. The median Anthropic employee's imply 17.1%, and the median visitor to the website explorer's 12.5%. The model's reference scenarios are Modest at 2.4%, Substantial at 5.4% and Extreme at 15.4%. The public median therefore sits at the Substantial scenario, which the authors describe as more than twice the typical US growth rate, while the staff median lies beyond Extreme. At 17.1% annual growth, the economy would double roughly every 4.5 years.
The distributions overlap widely, so many members of the public hold views as bullish as the typical staff member, and some staff are more cautious than the median.
8. What do staff think about AI building AI?
Employees were also asked what share of the work of building a frontier model is done by AI. The median answer was 51% today, 80% by the end of 2027 and 95% by the end of 2030. Staff on research teams estimated 60% for today, against 53% for other respondents, and the variation inside teams was larger than the variation between them.
A footnote adds context: more than 80% of the code merged into Anthropic's codebase is written by Claude, but building a frontier model also requires research direction, experiment design and evaluation. The authors again present these as personal, internal estimates and not as company measurements or forecasts.
9. What are the limits of the findings, and what comes next?
The authors list several caveats. The public sample is an opt-in online panel that leaves out people who are offline. The staff sample is internal, unweighted and not representative of the wider workforce. Website visitors selected themselves. Staff and workers answered some questions in different formats, with a slider for one and ranges for the other, so midpoints were used to compare them. The model's outputs depend on its framework and on fixed parameters, and the direction of causation between beliefs and exposure is not established.
Because Anthropic develops the technology it surveys, readers should treat the staff numbers as the expectations of insiders at one frontier laboratory, not as a forecast confirmed by outside evidence.
The report's own conclusion is that the two groups agree on present-day capabilities but part ways on the pace of progress, the extent of adoption and the size of productivity gains. It suggests that public views may shift as AI use spreads, noting that people who already use AI at work hold opinions closer to those of staff.
No follow-up survey date was announced, but the September replication already shows modest movement, most visibly in the Nobel-level discovery question. For policymakers, the practical point is the spread between planning for unemployment near 8% and planning for a rate above 12%, along with the question of how quickly displaced workers can be placed in new roles.