Anthropic's earlier studies of how people use and regard AI drew on people who use Claude: usage data in the Economic Index, and interviews with Claude users through Anthropic Interviewer. This post, published by the Societal Impacts team on 12 June 2026, reports the first wave of a survey series of the general US public, which the post says "marks the first time we've spoken to the general public, allowing us to reach non-users of AI". It asks what Americans hope AI will do, what they fear from it, how they use it, and what they want from government and from AI companies. The headline findings are that curing disease was the hope most often ranked in the top three, job loss was the most common worry, a majority in every state supported government involvement in AI, and AI companies were the institution least trusted to make decisions about AI. The survey was fielded about six months before publication and describes opinion in late 2025.
The web post is the only document. It has no named authors, links no questionnaire, toplines or dataset, and puts its methods in an appendix at the end of the page. Most numbers appear both in the text and as printed labels on its charts; locations below give the post's section heading and, for charts, the chart's title.
How the survey was done
From "Method in brief" and "Appendix: Methodology":
- Population and mode. An online survey of the US "late-teen and adult internet population, age 16 and over", resident in the 50 states, the District of Columbia or Puerto Rico. YouGov conducted the fieldwork from its online panel between 1 November and 11 December 2025.
- Sample design. 52 parallel state and territory samples, targeting about 1,000 completes each (about 500 in Montana, North Dakota, South Dakota, Wyoming, Vermont and Puerto Rico), with quotas within each state on age, gender, education and race or ethnicity. The achieved sample is 51,993 completes, from 232 in Alaska to 1,902 in New York.
- Weighting. Weighted to the US resident population aged 16 and over on state, age, gender, education and race or ethnicity, with weights normalized to the sample size. All percentages are weighted.
- Margins of error. National ±0.6 percentage points at 95% confidence for a proportion of 50%; state margins from ±2.6 points (California, New York, Texas) to ±9.1 (Wyoming). The post does not say how these were computed for a quota-and-weighted panel sample, or whether they include a design effect.
- Definitions. Percentages use the full segment as denominator, so "don't know" and skipped answers stay in the base. "Worried" on the fears battery means the top four boxes of a five-point worry scale (the text puts it as any response of 2, "somewhat worried", or higher). "As good or better" on the capability battery means the top three boxes of a five-point scale. "Integrated users" report using AI at least once a day for work and at least once a day for personal purposes (unweighted n = 2,717). Party groups include leaners.
The post does not report a response or completion rate, the full wording of most questions, the order in which batteries were asked, or who analyzed the data. A note under the fears chart says the fear groupings came from a principal components analysis with oblique rotation, keeping three components. The survey is a single cross-section, and the post describes it as a baseline for later waves.
What it finds
Hopes
Respondents chose their top three hopes from a list of 17 ("What Americans hope AI will deliver"). Helping cure diseases such as cancer and Alzheimer's was chosen by 48%, helping people with disabilities by 36%, and three items by 23% each: technological progress, making life easier, and automating boring tasks (chart: What Americans hope AI will deliver; the summary at the top of the post names only the first two of the three as tied). The lowest were making people less lonely (4%) and being a therapist or counselor (3%).
Fears
Respondents saw 20 possible harms, flagged those that concerned them and rated their worry on a five-point scale ("What Americans fear"). The post notes that hopes (ranked) and fears (rated) were measured differently and "the numbers aren't comparable." Shares worried (chart: The hierarchy of AI fears): job loss 64%, "ruin our ability to think for ourselves" (which the post calls cognitive dependency) 56%, misinformation 52%, use by criminals 50%, surveillance 47%, and use by terrorists to make weapons 43%. Acting "of its own accord" was at 35% and going rogue to "end civilization" was lowest, at 27%. Only the first three items had a majority worried. From these the post concludes that Americans "tended to be more concerned with the misuse of AI than AI misalignment".
Job loss across groups
Worry about job loss was the top fear among Democrats (67%) and Republicans (62%), in households with and without children (59% and 66%), and in every state, from 71% in Iowa to 57% in Mississippi ("Patterns with job loss"). By education (chart: Job loss concern by education), it was 68% for postgraduate degrees, 66% for four-year degrees, 68% for some college or two-year degrees and 59% for high school or less. The post's text says concerns "rise with a respondent's education level" and puts the postgraduate gap over high school at "nearly 10 percentage points"; the chart shows a gap between high school or less and every college group, with no steady rise among the college groups.
People who use AI at work every day were less worried about job loss than those who never do, 54% against 70% ("Fear of job loss is heightened among those who use AI least"). The text describes the second group as people "who don't use AI at all", but the chart (Job loss concern by AI usage at work) groups respondents by how often they use AI at work, on a base of all respondents, so "never" there includes people without jobs and people who use AI only outside work. The chart's steps are 54%, 55% and 56% for daily, several times a week and weekly use, then 64%, 63% and 70% for less frequent use and never. The post offers explanations (skills that let people augment their work, or seeing AI's limitations) and says "Likely it's a combination of these and other factors"; these are the authors' suggestions, and the comparison is cross-sectional.
Perceived capability and wanting AI at work
For 14 workplace tasks, respondents rated how well AI could do each today and how much AI involvement they would want in their own job ("Perceived capabilities and acceptance of AI at work"). The share saying AI is as good as or better than most people ranged from 75% for research to 44% for service and support (text). The chart (Perceived AI capability versus personal acceptance) prints no values; read from it, the share not wanting AI involved now was about 43% for information gathering and research and about 66% for service and support. That category combines "I never want AI involved in this" with "I don't want AI involved in this now, but maybe in the future" (chart note). The post says acceptance "seems to move in lockstep with perceived capabilities"; the chart shows the two measures ordered roughly inversely across tasks, without a reported correlation.
Cognitive dependency
Respondents were asked how much disruption they would feel if AI became unavailable tomorrow ("Cognitive dependency is an anticipatory fear"). The chart (Worry about dependency versus reliance on AI today) prints shares of all respondents: 42% worried and minimal impact, 12% worried and significant inconvenience or worse, 24% not worried and minimal impact, 16% not worried and significant inconvenience or worse, with cells not summing to 100% because of non-response. The post summarizes this as about one fifth of the worried and about one third of the unworried expecting significant disruption. By our arithmetic, 12 of the 56% worried is 21%, and 16 of the 44% not worried is 36%; taking only those who answered the disruption question (the cells' row sums, 54% and 40%), the shares are 22% and 40%. The post calls the fear "mostly anticipatory". Educators were among the occupations most worried about dependency, "second only to" arts and design; the chart (Dependency concern, top and bottom occupations, employed respondents, n = 24,508) prints both at 61%, against 39% for construction at the bottom. Those using AI daily at work were less worried about dependency than those who never do (46% against 62%).
Integrated users
About 6% of Americans were integrated users ("What the heaviest AI users tell us"). They skewed young, male, urban, employed and college-educated (chart: Integrated users demographic profile, for example +19 points employed and +12 points male against the national average), and nearly two-thirds described themselves as early adopters, against 30% of the public. They were less worried than the public about every listed harm, which the post says "probably reflects differences in the outlook of early adopters". The post's description of them as possibly "where mainstream opinion is headed" is its own speculation.
Government, industry and trust
- Government involvement. 71% said government should be involved in the development and regulation of AI: 79% of Democrats, 68% of Republicans and 69% of Independents, and a majority in every state and territory surveyed, from 81% in the District of Columbia to 63% in Hawaii ("What Americans want from AI governance"). The chart (Support for government involvement in AI) words the item as the government playing "a role in influencing or legislating different aspects of AI", and prints larger gaps by education and income than by party: 84% with postgraduate degrees against 65% with high school or less, and 80% with household incomes of $100K or more against 68% under $50K.
- Domains. Of eight domains, only privacy (56%) and child safety (52%) had majority support for a more-than-minimal government role; liability was 49% (chart: Support for more-than-minimal government role, by domain and party). The text says national security had the narrowest partisan gap, "just three points"; the chart prints 53% for Democrats and 49% for Republicans, a four-point gap as rounded.
- What would reassure people. Ranking options for ensuring AI is built in humanity's interest, 47% put legal liability for harm in their top three and 44% put safety over growth, followed by independent watchdogs (29%) and slowing development for safety (27%) ("What Americans want from the industry"). The chart shows Republicans ranking both leading items slightly higher than Democrats.
- Trust. 15% trusted AI companies to make decisions about how AI is developed and used, below the federal government (20%), state and local government (19%) and international bodies (20%), and far below independent experts (43%) ("The trust deficit"). The chart also lists voters in a direct vote, at 27%, which the text does not mention.
- Integrated users and governance. Integrated users trusted every institution more and were less inclined to slow or stop AI, but supported government involvement at about the national rate (74% against 71%).
The post's overall reading is that "AI did not heavily divide Americans along typical partisan, geographic, or educational lines", with disagreement "largely only in the intensity of people's views."
How the post relates its results to Anthropic
The closing section ("Anthropic initiatives and what comes next") links the findings to two Anthropic policy frameworks announced shortly before: an Advanced AI Framework that proposes mandatory independent safety testing, transparency requirements and government authority to block or recall dangerous deployments, and an Economic Policy Framework on preparing for job displacement. It states that "The direction AI takes should not be set only by the companies building it." These are Anthropic's positions. The survey was commissioned and reported by a developer that has published policy proposals on the same subjects, and the post presents the results alongside them.
Limitations and open questions
The post's stated limitation is that the survey "captures what Americans believed about AI in late 2025" and is a baseline. From the methods given:
- The sample is an opt-in online panel with quotas and weights, described by the post as "nationally representative"; the population is the internet population aged 16 and over, so it includes 16 and 17 year olds and Puerto Rico, and excludes people offline.
- Without the questionnaire, the wording and order effects of the batteries cannot be checked. Only the two capability and preference questions are quoted in full.
- The integrated-user comparisons and the use-versus-worry comparisons are associations in one cross-section; the post's causal explanations are offered as possibilities.
- Several text summaries are looser than their charts (the three-way tie in hopes, the education gradient, the three-point partisan gap, "don't use AI at all" for a chart of use at work), as noted above. The post's reading that AI did not heavily divide Americans along educational lines sits beside a 19-point education gap in support for government involvement.
- No data are released, so outside analysts cannot check the weighting or recompute subgroup results.
Research questions
- How do later waves move? Comparing the same items across waves, with wording changes recorded, would show whether worry about job loss and trust in AI companies change as adoption grows.
- Do the stated margins hold for this panel design? Published design effects, or comparison with probability-sample surveys fielded at the same time, would show whether the ±0.6 point national and ±2.6 to ±9.1 point state margins understate uncertainty.
- Does daily use lower worry, or do less worried people use AI more? Following the same respondents across waves would separate the two readings of the 54% against 70% gap.
- How sensitive is the ranking of fears to the format? A split sample with a forced ranking would show whether "misuse over misalignment" survives a change of method.
- Does low trust in AI companies extend to named developers? Items naming individual companies would show whether the 15% reflects the industry or particular firms.
- How do survey and interview findings compare? Matching items between Public Record and Interviewer studies of Claude users would show how far user-based findings differ from the general public's.
How it connects
- What 81,000 people told us about the economics of AI (April 2026) is the analysis the post cites as reflecting the finding that more exposed workers worry more about displacement. That study coded open-ended interviews with self-selected Claude users using Claude classifiers, a different population and method from this weighted survey.
- Introducing Anthropic Interviewer (December 2025) describes the interview tool behind the 81,000-person study the post cites for job loss and dependency fears.
- What do you want from AI? (September 2026) launches a second Interviewer study of Claude users with an option to publish interviews; it does not cite this survey, and the connection is this note's.
- The Anthropic Economic Index (2025) is the usage-data series the post names as related work.
- Preparing for AI's economic impact (October 2025) records the June 2026 Economic Policy Framework that this post cites.
References
- Anthropic (2026). Results from the first Anthropic Public Record. Societal Impacts, 12 June 2026. https://www.anthropic.com/research/anthropic-public-record (live page checked 13 November 2026 against the Wayback capture of 16 July 2026; the text and chart images are the same) record
- Huang, S., et al. (2026). What 81,000 People Want from AI. Anthropic, 18 March 2026. https://www.anthropic.com/features/81k-interviews record · 2 pages
- Anthropic (2026). Policy on the AI Exponential. https://www.anthropic.com/policy-on-the-ai-exponential record · 2 pages