{"id":574129,"date":"2026-09-30T13:55:19","date_gmt":"2026-09-30T17:55:19","guid":{"rendered":"https:\/\/alpha.pewresearch.org\/pewresearch-org\/?p=574129"},"modified":"2026-09-30T13:55:19","modified_gmt":"2026-09-30T17:55:19","slug":"how-synthetic-respondents-express-certainty-and-factual-knowledge","status":"publish","type":"post","link":"https:\/\/alpha.pewresearch.org\/pewresearch-org\/data-labs\/2026\/09\/30\/how-synthetic-respondents-express-certainty-and-factual-knowledge\/","title":{"rendered":"How synthetic respondents express certainty and factual knowledge"},"content":{"rendered":"\n
About this research<\/summary>\n

This Pew Research Center report examines whether AI-generated survey data can accurately duplicate the results of high-quality public opinion polls.<\/p>\n\n\n\n

We conducted this methodological experiment to better understand an emerging method in survey research. The Center believes that speaking to the public is essential to measuring public opinion and has no current or future plans to use AI models to generate survey results. Read more about our AI policy<\/a>.<\/p>\n\n\n\n

For this study, we had an AI model take on the role of real respondents from our American Trends Panel<\/a> (ATP) and answer the same surveys administered to human panelists. These surveys included many different question formats and asked about a wide range of topics, attitudes and behaviors.<\/p>\n\n\n\n

Why did we do this?<\/h4>\n\n\n\n

The Center regularly studies and evaluates advances and trends in polling research<\/a>. As AI-based polling becomes more widely used in the industry<\/a>, we wanted to understand the data these \u201csynthetic samples\u201d produce and how it compares with high-quality surveys of real humans on topics of broad public interest. That question is the focus of this study; it does not address other uses of AI models in polling.<\/p>\n\n\n\n

How did we do this?<\/h4>\n\n\n\n

We used a \u201cdigital twins\u201d approach for this study, asking an AI model to adopt the personas of real humans who are members of the ATP. We gave the model a wide range of information about each panelist<\/a>, including their self-reported demographic information and their responses to questions on our 2025 political typology survey<\/a>. We then showed the model questions from three different ATP surveys conducted in the first half of 2026 and asked it to respond to those same questions as the assigned persona. The model was given each question on the survey in order from start to finish, with exactly the same instructions and survey programming that the real humans who took it received.<\/p>\n\n\n\n

To make a direct comparison between our AI and human polls, we only replicated surveys for the ATP panelists from each wave who had also completed the political typology survey. As a result, figures for U.S. adults in this report may differ slightly from those previously published by the Center.<\/p>\n\n\n\n

We tested multiple models, but all synthetic results unless noted otherwise come from Anthropic\u2019s Claude Opus 4.6. At the time we conducted this study, Opus 4.6 was the most recent Anthropic model available for commercial use and had the best performance of several we evaluated.<\/p>\n\n\n\n

Here are the survey questions from Wave 185<\/a>, Wave 190<\/a> and Wave 192<\/a> used for this analysis, the detailed responses<\/a> from the ATP and AI surveys, and the methodologies<\/a> for the original ATP survey waves and our AI surveys.<\/p>\n<\/details>\n\n\n\n

Public opinion polling often highlights the extent to which people are unsure about an issue. Knowing and quantifying that ambiguity is essential to any understanding of how important public debates may be evolving.<\/p>\n\n\n\n

Similarly, knowledge questions in polls can help provide valuable insights into the public\u2019s familiarity with a topic \u2013 for example, whether people generally understand the digital privacy landscape<\/a> or know key facts about the U.S. system of governance<\/a>.<\/p>\n\n\n\n

In contrast to humans, large language models are armed with vast quantities of training data drawn from around the internet. As such, there is reason to think that synthetic surveys will overstate the level of factual knowledge the typical American possesses.<\/p>\n\n\n\n

These models also have a known tendency to want to provide answers<\/a> to questions and will confidently provide incorrect answers<\/a>. Because of this, we also might expect them to be consistently less willing than humans to say \u201cI\u2019m not sure\u201d on surveys.<\/p>\n\n\n\n

In our examination of synthetic polling, we found ample evidence that AI models do exactly these things.<\/p>\n\n\n\n

This analysis is part of a larger evaluation of AI-generated synthetic samples in public opinion research. Read<\/em> a summary of the main findings<\/a><\/em> and refer to the<\/em> methodology<\/a><\/em> for more details on how we conducted our synthetic poll and compared it with real survey results.<\/em><\/p>\n\n\n\n

\u2018Not sure\u2019 response options<\/h3>\n\n\n\n

We provide human survey-takers on the American Trends Panel<\/a> (ATP) with an explicit \u201cnot sure\u201d option on many questions we ask. Depending on how the questions are designed, this response may indicate uncertainty or indecision, or a lack of factual knowledge on a given topic.<\/p>\n\n\n\n

In the instructions to our synthetic survey-takers, we made clear that it was realistic for them not to know about certain topics or to be incorrect at times about factual matters. But despite this prompting, our synthetic respondents chose \u201cnot sure\u201d options far less often than real humans do.<\/p>\n\n\n\n

Across all opinion questions where \u201cnot sure\u201d was provided as an explicit option:<\/p>\n\n\n\n