# The AI expectations gap - source note Accessed: 2026-10-19. Publication: 2026-10-16, verified against the official PDF cover. Evidence label: technical report (survey research); respondent forecasts and conditional economic scenarios. Publisher: The Anthropic Institute. Authors: Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek and Peter McCrory. ## Primary sources - Official report: https://www.anthropic.com/research/ai-expectations-gap - Official PDF: https://www-cdn.anthropic.com/a28f24814c56e2834a3eae7b13946f19d3e97ec4/The%20AI%20Expectations%20Gap.pdf - Appendix linked by the report (not independently analyzed for this note): https://www-cdn.anthropic.com/1bc1a3067f9c9883d6d556b38f26fd4f1e1b640f/The%20AI%20Expectations%20Gap%20-%20Appendix.pdf The main report HTML and 16-page PDF were retrieved. PDF page numbers below count the cover as page 1. The public article links to the publisher's PDF; no redistributable-license assumption is made. Local research captures and hashes are retained outside the public deployment. ## Claim-to-source map - PDF p. 1: title, authors and publication date. The web page's suggested BibTeX says September while the PDF says October and its cover explicitly says October 16. Use the PDF date and preserve this discrepancy rather than silently treating both as identical. - PDF pp. 2-3: 39% of US adults versus 91% of Anthropic staff expect a Nobel-level discovery by 2030. Quote: "39% of US adults and 91% of staff expect that AI will be able to make a Nobel-level discovery by 2030." These are respondents' beliefs, not calibrated probabilities or observed achievements. - PDF p. 7 and Figure 4: 26% of working AI users versus 59% of staff report a suitable task takes a quarter of the time or less. Caption explicitly includes time spent checking and fixing work. Self-report, not randomized productivity evidence. - PDF p. 8 and Figure 5: average expected automatable share of current work hours is 28% for employed US respondents and 75% for staff. Staff used a slider; public respondents chose ranges and authors used midpoints. Do not call this a forecast that 75% of jobs vanish. - PDF pp. 9-10 and Figure 6: association between expected own-job exposure and unemployment forecasts. Association among beliefs is not causal labor-market evidence. - PDF p. 10: 95% of staff expect headcount growth. Quote: "These are the personal expectations of staff members in an internal poll and do not represent Anthropic's plans or its view of its headcount." - PDF pp. 11-12 and Figure 7: 5.4% and 17.1% are median model-implied GDP growth rates in 2030. Complete-input subsets: 3,259 US adults and 1,093 staff. Not cumulative GDP changes and not direct GDP poll answers. Economic model not independently reproduced. - PDF p. 15, footnote 1: Morning Consult public survey August 11-23, 10,980 adults, online opt-in, demographically weighted; split future horizons 2027/2030. Staff poll August 9-12, 1,334 responses, unweighted, all asked 2030. Total sample is not the denominator for every result. - PDF p. 15, footnote 2: September 17-24 repeat poll reached 7,613 adults; Nobel-level-discovery belief increased to 44%; main report still uses August results. - PDF p. 15, footnote 5: other economic-model parameters held at the "Substantial" scenario values. ## Editorial boundary This is evidence of an expectations gap, not proof which group is right. Distinguish capability, adoption, productivity, work hours and employment. Information advantage, task composition and selection are possible explanations; the survey does not identify their causal contributions. US results are not a Canadian jobs forecast. Do not imply peer review, independent replication, company hiring commitments or guaranteed growth. Practical workflow testing recommendations are our interpretation. ## Weekly source checks and selection Primary pages opened October 19: - Anthropic research: https://www.anthropic.com/research - extraction exposed team descriptions rather than a complete recent-item list; supplemented with domain-specific search and the full selected report. - Transformer Circuits: https://transformer-circuits.pub/ - displayed August interference-weights and July workspace entries; no newer item selected. - OpenAI publications: https://openai.com/research/index/publication/ - October system-card update and mathematics report visible; related mathematics/model coverage already exists in local research. No additional article in this run. - Google DeepMind research: https://deepmind.google/research/ - scientific-AI and publication listings checked; index extracts varied in recency, so not treated as exhaustive. - Google Research: https://research.google/blog/ and https://research.google/blog/2026/10/ - root extract appeared stale; October archive confirmed current listings including wet-lab agents and Deep Review. Leads retained, no detailed scientific conclusions published from these index entries. - Meta: https://ai.meta.com/research/publications/autoformbench-a-lean-formalization-benchmark-for-advanced-mathematics/ - official October 13 abstract checked; benchmark concerns building missing Lean infrastructure. Not selected for this week's single explanatory article. - LawZero: https://lawzero.org/en/blog - Scientist AI/implicit-agency/sycophancy material checked; no fresh dated October item verified in this extract. - DAIR: https://www.dair-institute.org/blog/ - long-form critique index checked; undated extracted headlines do not establish a new release. Search results additionally surfaced quantum-circuit and distillation-defense reports, and a DeepMind on-device scam-detection announcement. These are discovery leads only here, not independently analyzed findings in the published article. This is a bounded weekly scan, not certification of every named researcher's output. ## Duplicate and publication scope Before drafting, searched existing library, article HTML and top-level AI source notes for the exact report URL/slug; no match. This is distinct from the existing Economic Index Cadences note: that covers use telemetry, whereas this report compares public and staff expectations and maps beliefs to economic scenarios. One new explanatory article and one library card are prepared for this run. Also refresh the seed JSON, library count/latest links, AI entry link, blog index, search entry and sitemap. Deployment is limited to exact changed paths on Hostinger shared hosting, after remote baseline comparison and backup. Live verification is recorded separately; this note itself is not evidence that deployment succeeded.