Bottom line: a survey is not a crystal ball
Anthropic's new report finds a large gap between what its staff and US adults expect AI to do by 2030. It is useful evidence about beliefs at a frontier laboratory and beyond it—not proof that either group has correctly predicted the future. Expectations about automating tasks are also not the same thing as forecasts of how many jobs will disappear.
What was actually studied
The AI Expectations Gap: Public and Frontier Lab Views on AI's Impact on Work, by Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek and Peter McCrory, is an Anthropic Institute technical report. The official PDF is dated October 16, 2026. Its main comparisons use surveys conducted in August—not measurements of what AI can already do in October.
Morning Consult surveyed 10,980 US adults through an online opt-in panel between August 11 and 23. Responses were weighted to represent the adult population, but the authors acknowledge that demographic weighting cannot remove every selection difference. Half received questions about 2027 and half about 2030. Anthropic's separate August 9–12 internal poll drew 1,334 staff responses, was unweighted, and asked everyone about 2030. These are different populations, with different jobs and everyday exposure to AI.
The gap is real; the future outcomes are not yet known
- Scientific capability: 39% of US adults and 91% of surveyed Anthropic staff expected AI to be capable of a Nobel-level discovery by 2030. Those percentages measure respondents' expectations, not the probability established by a scientific test.
- Their own work: employed US respondents expected AI to be able to perform an average of 28% of their current work hours by 2030. Anthropic staff put that figure at 75%. This is a share of work hours, not a share of workers who will lose their jobs.
- Perceived time savings: for a task well suited to today's AI, 26% of US adults who use AI at work and 59% of staff said it takes a quarter of the time or less with AI. The question included checking and fixing the output. These are survey answers, not a controlled productivity experiment.
The authors also report a September repeat survey of 7,613 US adults. The Nobel-level-discovery expectation rose to 44%; the main analysis nevertheless uses August throughout. Keeping the survey wave attached to the number matters.
Why proximity to a lab is informative—but not decisive
People building frontier systems may see capabilities earlier and use them more intensively. That makes their perspective worth studying. It does not make their workplace representative of a hospital, construction company, school or small business. The report itself finds that regular AI users outside Anthropic tend to have expectations closer to staff expectations, but not identical to them.
Several explanations remain possible: different information, different tasks, different adoption experience and selection into a technology company. The survey does not isolate which explanation causes the gap. It therefore cannot tell us simply that the public is uninformed, or that staff are overconfident. Both shortcuts go beyond the evidence.
How a survey becomes a dramatic GDP scenario
The report maps respondents' beliefs about capabilities, adoption, automation, productivity and worker transitions into an economic model. For respondents who supplied all required inputs, the reported median model-implied US GDP growth rate in 2030 is 5.4% for the public sample and 17.1% for staff. The relevant subsets are 3,259 US adults and 1,093 staff—not the full survey samples.
These are conditional model outputs. They are neither observed growth nor direct answers to a GDP forecast question. They combine survey beliefs with the model's structure and additional assumptions; the authors hold other parameters at their “Substantial” scenario settings. The figures refer to growth in 2030, not cumulative growth from now to 2030. We have checked the report's description, not independently reproduced its economic model.
Automating a task does not settle the employment question
Despite expecting extensive automation of their own tasks, 95% of surveyed staff expected Anthropic's headcount to grow over the following year. The report explicitly says these are employees' personal expectations, not company hiring plans. The combination is a useful reminder: a business can automate tasks and expand output at the same time.
That does not guarantee a painless transition elsewhere. In both groups, respondents who expected more of their own work to be automatable also tended to forecast higher unemployment. This is an association between beliefs, not a demonstrated causal estimate of future job losses. Nor is a US survey a forecast for Canadian employment.
What a useful next test would look like
Our interpretation is that organizations should treat this report as a reason to test assumptions, not as a mandate for either layoffs or complacency. Measure completed work, correction time, error rates, customer outcomes and costs in real workflows. Follow whether claimed time savings persist outside highly AI-intensive firms. Keep capability, adoption, productivity and employment separate in the ledger.
Repeated surveys can show how beliefs change. Independent workplace studies and observed labor-market outcomes are needed to establish what actually changes. The value of this paper is making an expectations gap visible; its limitation is that expectations remain expectations.
Primary sources and evidence trail
Evidence label: technical report—survey research; future outcomes are respondent forecasts and conditional economic scenarios. This note does not claim peer review or independent replication.
- Official report page
- Official PDF, October 16, 2026—findings pp. 2–3, time savings p. 7, work hours p. 8, hiring p. 10, model scenarios pp. 11–12, survey methods p. 15.
- Dated source note, method caveats and weekly source checks