The data are responses from an Epoch AI/Ipsos survey fielded on Ipsos’ KnowledgePanel, a probability-based online panel that uses address-based sampling to choose respondents. For the analysis, we use three waves of sampling: March 3-5, March 13-15, and April 3-5, 2026. Among other questions, the survey asks respondents which AI services they used in the past week. All estimates are weighted to be representative of the adult U.S. population. Household income is drawn from KnowledgePanel profile data. The number of respondents are 2,021 in wave 1; 2,028 in wave 2; and 1,020 in wave 3.

For this analysis, we focus only on the question about usage of specific AI tools in the past week and and reported household income from respondents’ profile data. We also remove a small set of clearly invalid responses. In all three waves, respondents who reported using the fictitious service “Clarity AI” are excluded. For wave 1, we also exclude respondents who reported a Clarity AI subscription. This removes 25 observations in wave 1, 17 in wave 2, and 1 in wave 3.

We compute baseline US adults’ household income numbers from 2024 CPS ASEC (Census HINC-03): share of adults living in households at each income level, with ‘children of householder’ excluded to approximate the 18+ population. On this basis, 50% of US adults live in households earning $100,000 or more, and 24% live in households earning under $50,000.

Responses from all three survey waves are pooled before computing income shares to maximize sample sizes. Pooled sample sizes for weekly users: ChatGPT (n = 1,611), Claude (n = 201), Google Gemini (n = 1,105), Grok (n = 221), Meta AI (n = 359), and Microsoft Copilot (n = 590). Users may report using more than one AI service, so these groups are not mutually exclusive.

For each service, we compute weighted income-bracket shares within that service’s weekly users. Specifically, we divide the sum of survey weights for respondents in each income bracket by the total sum of survey weights for all of that service’s weekly users. This yields a percentage breakdown that accounts for the probability-weighted representation of each income group.
To quantify uncertainty, we compute weighted standard errors using Taylor series linearization (following our polling hub methodology). We then construct 90% confidence intervals using these standard errors.

From the graph, we exclude services with smaller shares (Perplexity, DeepSeek, and Character.AI, ‘other’ responses, and non-responses.

Each wave surveys a fresh sample (cross-section) rather than tracking the same individuals, so results reflect population-level changes. Pooling waves increases sample sizes and requires that the income profile of each service’s users be roughly stable across the three survey field periods (March, April 2026). In a series of Rao-Scott survey-weighted tests for homogeneity, we find no statistically significant evidence that any tool’s user income profile differs across waves.

Claude (n = 201) and Grok (n = 221) have relatively small pooled samples among the six charted services, so income-bracket estimates for these services carry wider uncertainty than for other services.

Although results are weighted to be representative, as with all surveys, non-respondents may differ from respondents in ways the weighting does not capture. Usage is self-reported and may be affected by recall error or misclassification.