The data are responses from the Epoch AI/Ipsos survey fielded on Ipsos’ KnowledgePanel, a probability-based online panel that uses address-based sampling to choose respondents. 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. 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. We also remove a small set of clearly invalid responses. In all three waves, respondents who selected the fictitious service “Clarity AI” are excluded. In 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.

For each wave, we estimate Claude usage as a percentage of all US adults — specifically, the weighted proportion of adults who reported using Claude in the past week. This produces three plotted points, with point estimates of 3.0% in wave 1, 4.4% in wave 2, and 4.3% in wave 3.

To quantify uncertainty, we compute weighted standard errors using Taylor series linearization (see our polling hub methodology). We then construct 90% confidence intervals using a logit transformation. While the simple Wald interval on our polling hub is sufficient for most exploratory analysis, Claude’s estimated share was small enough to merit the logit transformation.

To test for an upward trend, we fit a weighted linear trend to the log-odds of the median dates of each of the three waves. We estimate using weighted least squares, with each wave weighted by the inverse variance of the logit of the estimated proportion at the time. In effect, this gives more influence to waves which are measured more precisely. We test the one-sided null hypothesis of the trend being flat or downward. Under this specification, the slope of Claude’s user share is positive and statistically significant at the alpha=0.05 level on a one-sided test (p = 0.0448). No other tool is significant at this level under the same trend test.

If one instead treated the analysis as ten simultaneous hypothesis tests (instead of testing the hypothesis of a Claude upward trend, and then testing the other services separately) and applied Benjamini-Hochberg or Bonferroni correction, none of the service trends would remain significant (including Claude). We mention this point as a robustness check, but our primary inference remains the single pre-specified Claude trend test.

Each wave surveys a fresh sample (cross-section) rather than tracking the same individuals, so results reflect population-level changes. The share of respondents reporting Claude use is small (roughly 50–90 respondents per wave), so estimates are subject to substantial sampling variability despite statistical significance.

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.

The trend is based on three points in time over approximately one month, so it may reflect a short-term spike rather than a sustained shift. The survey does not capture why people use a particular service, so the observed increase cannot be causally attributed to any specific event.