About the Data

Responses are weighted to match U.S. Census targets on the joint distribution of gender, age, race, and education (96 cells). The unweighted sample is 2,122 respondents. The footnotes beneath charts show the Kish effective n (marked with ≈), which accounts for how unequal the weights are and so tends to be smaller than the unweighted total. The per-group labels shown on the split charts are unweighted counts (marked with =). Confidence intervals are Wilson score intervals computed on the effective n.

We generally report percentages out of everyone who was asked that question. So "Not Sure" is counted in the total and shown as its own segment on each chart. A figure like "about half of Americans expect AI to increase inequality" therefore means about half of everyone surveyed, not half of those who took a position. The one exception is questions that did not apply to a respondent: when people were asked about which tasks they would trust AI to perform, the response interface offered "Not Applicable," and a few workplace questions offered "I don't use AI." Those replies are left out of both the numerator and the denominator, because the question was not relevant to that person. Finally, only people who were working were asked questions about their workplace. The question-by-question view notes how many such replies were set aside for each question.

The groups come from a Latent Class Analysis (LCA), a clustering method that finds groups of respondents with similar answer patterns. Fit on 26 response items, the model favored six groups (K = 6) on BIC. We checked this clustering solution by measuring cross-validated log-likelihood and split-half stability. The LCA was fit with stepmix, using its categorical_nan option (full-information maximum likelihood) so respondents with partial answers contribute everything they did answer. The model is fit by expectation–maximization.

The quiz scores your answers on 12 of the 26 questions, matches them against each group's response pattern, and reports your closest match. All scoring happens privately in your browser.

Three factors emerged from an exploratory factor analysis of weighted polychoric correlations. Factors were extracted with the factor_analyzer package, using oblique (oblimin) rotation so any correlation between the factors would have shown up. Each respondent's score on a factor is a weighted combination of the items below, rescaled to the 0–1 range. For the map page, state-level "Trust in AI" / "Support for AI regulation" / "Workplace AI dependence" on the map are state means of these per-respondent scores.

Trust in AI capability. This factor loads primarily on questions about what a person would trust AI to perform: how much respondents would trust AI to look up factual information, file their taxes, manage their retirement savings, provide medical advice, or decide their court case.

Support for AI regulation. This factor loads on disclosure and accountability items: should companies be required to disclose to users when they are talking to an AI, to label AI-generated images and video, to inform applicants when AI is used in hiring. Should impersonation of a real person's voice or face be illegal? Should you have a right to a human review of an AI decision that goes against you?

Workplace AI dependence. This factor loads on two items: whether respondents feel pressure to use AI tools to keep up at work, and how much losing access to AI would affect their ability to do their work.

In practice, trust and regulation were nearly independent: Americans who trust AI are about as likely to want it regulated as those who do not. The full correlation map shows the three factor blocks along the diagonal.

Mode and language. The survey was administered online in English. Adults without internet access, and adults whose primary language is not English, are under-represented relative to the U.S. adult population, even after weighting on demographics.

Self-report. Frequency of AI use, workplace pressure, and prior exposure are self-reported. Respondents may over- or under-report use depending on what they understand "AI" to mean.

Response clustering and assignment. When we clustered the survey data using LCA, the 5, 6, and 7 cluster solutions all fit the data about equally well. We reported K = 6 because it minimized the BIC and produced interpretable groupings, but the clustering is not definitive. When we use the full set of 26 questions to assign participants to groups, the average respondent's estimated posterior probability of belonging to their top class is 0.93, and 86% have a top-class probability above 0.8; a small minority sit near a boundary and could plausibly belong to a neighboring group.

When we use only the 12-question quiz data, our certainty is lower: the quiz questions reproduce the 26-item full-model group assignment for 87% of respondents, and the full-model group is in the quiz's top two for 98% of respondents.

Geographic granularity. Sample sizes are small for most U.S. states (~42 per state). Therefore, state-level numbers should be read as suggestive rather than definitive and the geographic variation shown on the map should be interpreted with caution.