The AI expectations gap

An Anthropic Institute survey report (October 2026) comparing 10,980 weighted US adults with 1,334 Anthropic staff in August 2026; staff expect faster capability gains, wider use and larger time savings, expect AI to do 75% of their own work by 2030 against 28% for US workers, and imply 17.1% GDP growth in 2030 against 5.4% through an economic model, while in both groups unemployment forecasts rise with expected exposure of one's own work.

Szymon Sacher et al. (The Anthropic Institute) · 2026-10-16 · original

Forecasts of AI's economic effects vary widely. This report, published on 16 October 2026 by Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek and Peter McCrory for the Anthropic Institute, asks where the disagreement lies: in beliefs about how fast capabilities will improve, or in beliefs about how fast and in what form they will be adopted ("Introduction"). It compares a weighted national sample of US adults with a poll of Anthropic staff, people who "build and use frontier AI every day", on the same questions, and runs each respondent's answers through an economic model to see what growth they imply. Its full title in the suggested citation and the PDF is "The AI Expectations Gap: Public and Frontier Lab Views on AI's Impact on Work".

The report was released with How employment is changing at Anthropic, a study of Anthropic's own workforce, and a short post presenting both, Working at the frontier shapes expectations for the future. It is a developer's survey report without peer review; the response data are not released.

Version of record. The web page and its 16-page PDF carry the same text and eight figures. A separate 6-page appendix PDF gives the sample details, a decomposition of the growth gap (Figure A1), the full time-savings answers (Figure A2) and two external checks against the Anthropic Economic Index (Figures A3 and A4). The model used to translate answers into growth is in a separate working paper by the same five authors, Economic Scenarios for Transformative AI (Korinek et al., Working Paper 2026-02, September 2026); this note draws on its survey sections (3.5 and 4.4) only. That paper already reported the same August Morning Consult sample of 10,980 US adults and its mapping into the model (its Table 2 and Table 3), so what is new in this report is chiefly the staff poll, the comparison and the website quiz.

What the report does

Samples. Morning Consult fielded the public survey online, to opt-in panelists, from 11 to 23 August 2026. It collected 16,395 completed interviews; after keeping only each person's first interview, 10,980 remained. Each of the eight daily samples had been weighted to the US adult population, and after the repeats were removed the weights were recomputed so that the sample reproduces the full weighted sample by age and gender, race and ethnicity, education, census region and household income (Appendix A). The authors note that an online opt-in panel excludes adults without internet access "and may differ from the population in ways weighting on demographics cannot correct" (footnote 1). Several forward-looking questions were asked about 2027 to a random half and about 2030 to the other half; the report gives 2030 answers unless it says otherwise. The staff poll was internal, ran from 9 to 12 August 2026, drew 1,334 respondents from every department, is unweighted, and asked everyone about 2030. The report does not give the staff response rate or Anthropic's headcount at the time.

Response formats differ between the groups. The public could answer "not sure" (chosen by 12% to 21% depending on the question), and those answers are dropped question by question; staff had no such option but could skip questions (Appendix A). Some questions had ranges for the public and a slider for staff: for the share of one's own work AI could do, public answers are converted to range midpoints (Figure 5 note), and for re-employment time, staff answered in months on a slider (footnote 4).

Questions. The surveys covered when AI will be able to do each of eight tasks "as well as a skilled professional", from routine emails to a Nobel-level discovery (Figure 1); what share of feasible tasks will be done with AI by 2030; who will do five tasks by 2030 (AI alone, AI with a worker checking, a worker with AI helping, a worker alone; Figure 3); how long a task suited to AI takes with it, counting checking and fixing (Figure 4); what share of one's own current work hours AI could do by 2030 (Figure 5); a 2030 unemployment forecast, after being told the current rate was about 4.2% (Figure 6); and how long a worker displaced by AI in 2030 and needing a new occupation would take to find a job. That question's wording is printed in footnote 4; it begins "In normal times, someone who loses a job typically takes about 3 months to find a new one." The public alone was asked about obstacles to adoption (Figure 2); staff alone were asked about Anthropic's headcount and about AI's share of the work of building a frontier model (Figure 8).

From answers to growth. Answers on five dimensions (capabilities, adoption, autonomy, time savings and re-employment) are mapped, following Appendix B of the working paper, into inputs for its economic model, with every other parameter held at the paper's "Substantial" scenario values (footnote 5). This is done for each respondent who answered all five questions: 3,259 US adults (2030 half only) and 1,093 staff. The implied 2030 GDP growth rates are model outputs from stated expectations, not forecasts made by respondents.

A second wave. Morning Consult repeated the questions from 17 to 24 September 2026 with 7,613 US adults. The authors write that the main results "changed little": the median implied growth was 5.6%, and the share expecting a Nobel-level discovery by 2030 rose from 39% to 44% (footnote 2). The report uses the August results throughout.

What it finds

Capabilities and adoption

  • Today versus the future. The two groups rank the eight tasks in nearly the same order of difficulty and "largely agree" on what AI can do now (Figure 1, panel A). About a quarter of each group say AI can already design, conduct and publish scientific studies; for 2027, 9 in 10 staff expect this against 38% of the public. By 2030, 39% of US adults and 91% of staff expect AI to be able to make a Nobel-level discovery. Between 19% and 40% of US adults, depending on the task, say AI will never be able to do it, against at most 2% of staff ("Capabilities").
  • Use at work. Among US adults who use AI at work regularly (about a fifth of respondents), 56% expect AI to manage at least seven of the eight tasks by 2030, against 41% of all US adults, 28% of those who have never used an AI tool, and 96% of staff.
  • Adoption. Of the work tasks AI will be able to do, US adults expect 42% to be done with AI by 2030 and staff 65% ("Adoption and its obstacles").
  • Obstacles (public only). Choosing up to three of nine, US adults most often named customers preferring to deal with a person (41%), then data privacy and security (37%), AI not being reliable or accurate enough (27%) and laws, regulations or liability (27%) (Figure 2, which prints the shares). Regular users at work put privacy first (42% against 37% for others) and customer preference lower (37% against 42%).
  • Autonomy. Most staff expect AI by 2030 to do the five tasks alone or with a worker checking; about 4 in 10 US adults expect humans to do most of the work themselves or with some AI help (Figure 3).

Time savings

85% of staff say a task suited to AI takes half the time or less with it. Among US adults who use AI at work the typical answer is half the time, and 29% say it takes as long or longer, against 5% of staff. A quarter of the time or less is reported by 26% of US adults who use AI at work, 21% of those who do not, and 59% of staff (Figure 4; full distributions in Figure A2). The authors suggest that staff "may be describing what they experience at work", citing the companion employment report's finding that the code engineers added "tripled in the month after adopting coding agents" ("Time savings"). The employment report itself says the code added "rises sharply" and plots the estimate in log points without printing values; its month-0 coefficient is about 1.0 (read from its Figure 4), and e to the power 1 is about 2.7 (this note's arithmetic), so "tripled" is a rounded reading of that plot.

Own work and the labor market

  • Own work. On average, staff say AI will be able to do 75% of their current work hours by 2030, and US workers (respondents with a job) 28%; 1 in 20 US workers say 80% or more (Figure 5). Figure 6, panel B, prints the share of each group in each band: 55% of staff and 5% of US workers chose 80% or more. Among US workers the average is 25% for manual trades, production and transport, and 33% for computer, engineering and science occupations; regular AI users in that last group say 40%.
  • Unemployment forecasts. US workers who expect AI to do less than half of their work forecast 2030 unemployment of 7% to 8%; above half, forecasts rise with expected exposure, to an average of 12.2% for those at 80% or more (Figure 6, panel A). Across US workers, ten more percentage points of expected exposure is associated with 0.4 points more forecast unemployment. Holding exposure fixed, how many of the eight tasks a worker expects AI to master is "unrelated to the forecast" ("Own-job exposure and unemployment forecasts"; no coefficient or test is printed for either statement).
  • Same gradient in both groups. Staff forecasts follow a similar rise with exposure (Figure 6, panel A, which omits groups with fewer than 20 respondents and has no staff point below 20% exposure). From this the authors conclude that the groups "do not seem to hold different theories" of how AI affects their own jobs and the labor market, but have "different experiences of what AI capabilities are". The comparison is of group means with 95% intervals; the report does not state a formal test of equal slopes. The text says that at the same level of expected automation "both groups forecast about the same unemployment rate"; in the figure, staff means lie above US workers' in each band where both appear, by about 1 point at 20% to 34% and about 2 points at 35% to 49%, where the two intervals do not overlap (read from Figure 6, panel A, which prints no values for these points).
  • Hiring and re-employment. 95% of staff expect Anthropic's headcount to grow over the next 12 months, the median by 66%, and 27% expect it at least to double; the report states these "do not represent Anthropic's plans". The median staff member expects a displaced worker to need 12 months to find work in a new occupation; the median US adult answer is in the 7 to 12 month range ("Hiring and re-employment").

Implied growth

  • Medians. The median US adult's answers imply 5.4% GDP growth in 2030 and the median staff member's 17.1%, with wide, overlapping distributions (Figure 7, which shows the 10th to 90th and 25th to 75th percentiles). The working paper's Modest, Substantial and Extreme scenarios are 2.4%, 5.4% and 15.4% growth. The conclusion describes the staff median as GDP "doubling in size roughly every four and a half years".
  • Which beliefs drive the gap. Swapping staff answers into the US adult answers one dimension at a time, and averaging over all 120 orders, time savings add 4.9 points, adoption 3.6, capability 2.7, autonomy 0.8 and re-employment minus 0.2 (Appendix B, Figure A1, which prints these values and a total of 11.8 points; the order matters, so the figure also shows the smallest and largest contribution of each). The figure's subtitle reads "5.4% → 17.2%", against 17.1% in the text, a difference of 0.1 point that the report does not explain.
  • Website visitors. An interactive scenario explorer published on 9 September 2026 asked shorter versions of the questions; 23,075 visitors from more than 140 countries completed it in the first 48 hours, and their median implies 12.5% growth. The authors note these visitors chose to take the quiz and are "likely to be more interested in AI than the general public" ("Responses from online scenario explorer quiz").

Inside the lab

Staff were asked what share of the work of building an early-2026 frontier model AI does. The median answers are 51% today, 80% by the end of 2027 and 95% by the end of 2030 (Figure 8). Research-team staff put today's share at 60% on average, against 53% for others; the authors say this is "consistent with information playing some role" but also that researchers may use AI more in their own work. They label all staff answers "personal estimates" that "are not a measurement by Anthropic or a company forecast" (Figure 8 caption). Footnote 6 contrasts this with the statement in When AI builds itself that more than 80% of code merged into Anthropic's codebase is written by Claude.

External checks

Appendix D compares the survey with the Anthropic Economic Index. Across states, one unit of a state's Claude usage index goes with 8.1 more percentage points of adults using AI at work (standard error 2.0, correlation 0.57; Figure A3). Across the 45 countries with at least 50 quiz visitors, visitors per working-age person correlate with the country usage index at r = 0.87 in logs, while a country's median implied growth is unrelated to it; the median US visitor implies 12.2% growth and the median visitor elsewhere 12.7% (Figure A4).

The survey questions as written

The report has no principles or model prompts. What shapes its results is the wording and format of the questions. The report prints the re-employment question in full (footnote 4) and paraphrases the others in figure captions. Two features are stated in the report and matter for comparing groups: the public could answer "not sure" and staff could not, and the two groups answered some questions in different formats (ranges against sliders). The time-savings question asks respondents to count "the time spent checking and fixing its work" (Figure 4 caption). The report and its appendix do not link a full questionnaire, so question order and the exact wording of most items are not available.

Limitations and open questions

  • Two different samples. The public sample is weighted and drawn from an opt-in online panel; the staff poll is unweighted, of unknown response rate, and its respondents work for the company publishing the report and selling the product. Differences in format and in the "not sure" option complicate direct comparison, and the report does not test their effect.
  • Expectations, not outcomes. Every result is a stated expectation. The growth figures are model outputs that depend on the working paper's mapping and on holding other parameters at the Substantial scenario. The median US adult's implied growth (5.4%) equals that scenario's growth rate; the working paper, which describes the Substantial scenario as corresponding to forecasts that "circulated from financial and consulting institutions in 2023" (its Introduction), not as fitted to the survey, reports the same sample's answers as delivering "outcomes close to the substantial change scenario" (its Section 4.4).
  • Causal reading. The claim that experience drives the gap rests on comparisons by AI use within the public and between groups; the report gives no design that separates experience from selection into AI use or into working at Anthropic.
  • Staff views on headcount and AI's share of model building are labeled by the authors as personal estimates, not company plans or measurements.

No outside analysis or replication of the report was found for this note.

Research questions

  1. Does the gap persist when formats are matched? Staff used sliders and had no "not sure" option. Fielding the public questions to staff in identical format would show how much of the gap is format.
  2. Do expectations predict outcomes? Re-asking the same panel in 2027 and comparing with measured adoption, time savings and unemployment would show which group's expectations were closer.
  3. Does using AI at work change expectations, or do optimists choose to use it? A panel that follows the same respondents as they start using AI at work would separate experience from selection.
  4. How sensitive is the 5.4% to 17.1% gap to the model? Rerunning the mapping with parameters held at the Modest or Extreme scenario instead of Substantial would show how much the gap depends on that choice.
  5. Is the exposure-unemployment gradient the same slope in both groups? A pooled regression with a group interaction, reported with its interval, would test the "same theory" reading directly.
  6. Did the second wave change more than the Nobel item? Publishing the September wave's full results beside August's would show which expectations moved in a month.

How it connects

References