Changelog - August 14, 2026. Confidence intervals are now 90% Wilson score intervals computed on the effective sample size; the previous normal-approximation intervals (p ± 1.645×SE) could produce negative or zero-width bounds for small subgroups. Estimates use Ipsos’s final re-raked weights (changes of at most 0.2 percentage points; analyzed sample 1,103). Individual-level microdata has been added as a download (SPSS .sav and CSV).

This polling was conducted by Epoch AI in partnership with Ipsos on the KnowledgePanel, a probability-based online panel recruited via address-based sampling. Probability-based sampling reduces self-selection bias and produces estimates that are more representative of US adults.

The survey was fielded July 10-19, 2026, and includes 1,103 respondents. Respondents qualified for the survey only if they were employed or self-employed, full time or part time. A subset of the questions were asked only of participants who said they had used an AI service in the past week.

We create the ten-item task list using O*NET (version 30.3), which is the US Department of Labor’s occupational database. This database has become the shared task vocabulary in recent research on AI and work. We use intermediate work activities (IWA) from O*NET’s hierarchy. This 332-item layer gives enough detail while being common across various jobs.

We combined O*NET with national employment data to select tasks that are economically relevant. First, we ranked the O*NET activities by how common they are in the US workforce. To do this, we used national employment data from the BLS Occupational Employment and Wage Statistics (May 2025) to weight O*NET’s own task ratings for each occupation. To keep the list representative for various knowledge work jobs, we allowed up to two items per activity family. Medical activities were excluded.

We restricted the list to knowledge work activities, where AI services could plausibly be used. We excluded four items and replaced them with the next-ranked ones:

• executing financial transactions
• internal coordination/communicating about operational plans
• training others
• purchasing.

Within the remaining tasks, we did not select tasks by predicted AI exposure or current AI use. The task list was shown to respondents verbatim as O*NET labels.

The questions about individual work tasks were asked only of respondents who said the task is a regular part of their job. Follow-up questions on how long respondents have used AI for a task, how much time the task takes, and how much they edit the output were asked for at most five of a respondent’s AI-assisted tasks. If more than five tasks qualified, a random subset was selected, always retaining tasks where AI does most or all of the work. A randomly selected 10% of respondents who used AI for work were asked an open-ended version of the work-task question instead of the structured version.

The results have been weighted to represent the employed adult population of the United States, ages 18 and older, using US Census benchmarks. A small set of non-valid responses was removed before the analysis; the sample size of 1,103 reflects this removal.

The uncertainty intervals in the graphs are 90% Wilson score confidence intervals, computed on an effective sample size that accounts for unequal weighting. For each weighted proportion \(\hat{p}\) we compute

\(\frac{1}{1 + z^2 / n_\text{eff}} \left( \hat{p} + \frac{z^2}{2 n_\text{eff}} \pm z \sqrt{ \frac{\hat{p} (1 - \hat{p})}{n_\text{eff}} + \frac{z^2}{4 n_\text{eff}^2} } \right)\)

where \(z = 1.645\) is the 90% two-sided critical value from the standard normal distribution, and \(n_\text{eff}\) is the Kish effective sample size,

\(n_\text{eff} = \frac{\left( \sum_i w_i \right)^2}{\sum_i w_i^2}\)

computed over the respondents in the relevant base. Unlike the normal-approximation interval, the Wilson interval always lies within 0–100% and remains informative for proportions near 0% or 100%.

All estimates are weighted. For questions answered once per respondent, the estimated proportion for any response category is:

\(\hat{p} = \frac{\sum_i w_i y_i}{\sum_i w_i}\)

where \(w_i\) is the weight for respondent \(i\) and \(y_i\) is a binary indicator (1 if the respondent selected the category, 0 otherwise).

Standard errors, reported in the downloadable data, are computed using Taylor series linearization to account for weighting. The Ipsos KnowledgePanel uses an unclustered and unstratified probability-based sampling design. For questions answered once per respondent, the variance estimator therefore simplifies to:

\(\hat{V}(\hat{p}) = \frac{n}{n - 1} \cdot \frac{1}{(\sum_i w_i)^2} \cdot \sum_i \left[ w_i (y_i - \hat{p}) \right]^2\)

where \(n\) is the number of respondents in the relevant base.

Some questions are measured at the level of individual task reports, and one respondent can contribute several task reports. Because respondents, not task reports, are the sampled units, we treat the respondent as the primary sampling unit for these questions. The estimated proportion is:

\(\hat{p} = \frac{\sum_i w_i y_i}{\sum_i w_i d_i}\)

where \(y_i\) is the number of task reports from respondent \(i\) in the category and \(d_i\) is the number of valid task reports from respondent \(i\). The variance estimator is:

\(\hat{V}(\hat{p}) = \frac{n}{n - 1} \cdot \sum_i (z_i - \bar{z})^2\)

where \(z_i = w_i (y_i - \hat{p} d_i) / \sum_i w_i d_i\) and \(n\) is the number of respondents with at least one valid task report. For questions answered once per respondent (\(d_i = 1\)), this reduces to the variance estimator above. This accounts for the correlation between task reports from the same respondent. For these questions, we report the number of task reports and the number of respondents they come from. For these questions, the effective sample size is likewise computed at the respondent level, from the respondent weight totals, so that task reports from the same respondent are not treated as independent observations.