State of the Prompt

Reports / Working paper

Working paper · October 2026

The elusive verification: How people respond to AI information about politics

When ChatGPT answers a question about politics or government, what do people do next? Across a year of conversations, most answers drew no reaction at all. Pushback and requests for evidence were rare, even when the stakes were high.

Katherine Ognyanova, Mark Aakhus, Kiran Garimella, Shagun Jhaver, Eu Gene Lee, Jianing Li, Sarah Shugars

Rutgers University

Read the paper (PDF)

Key findings

  • 19%of political conversations contained any reaction to ChatGPT’s answer, though users wrote again in 72%. Most follow-ups were new requests.
  • 0.22%of political conversations asked ChatGPT for evidence, and 0.36% brought in an outside source. 1.6% disputed an answer.
  • 80%of disputes told ChatGPT what the user believed instead. Few cited a source, and most disputes ended unresolved.
  • 4.6×the odds of disputing an answer when the user stated a view before ChatGPT answered, holding other factors constant.

Abstract

Americans are routinely urged to be mindful of the information they encounter. Engaged consumers are expected to check sources, evaluate credibility, examine evidence, and think critically. With conversational AI emerging as a new information source, these expectations now extend to content we receive from chatbots. Yet we know little about people’s reactions to the answers that chatbots give them in everyday life. In this study, we analyze a year of ChatGPT logs from 73,852 members of an opt-in US consumer panel. We measure how users react to AI answers in 283,139 conversations about politics and government. We find that observed behavior fell short of normative expectations. Pushback and requests for evidence were rare, even when the stakes were high. When users did challenge an AI answer, they mostly did it by asserting views they already held rather than by seeking new evidence.

Read the full paper (PDF, 9 pages)

About the data

Political conversations were identified by a language model checked against human coding, and a second open model, run on Rutgers’ computing cluster, coded how users reacted to each answer. A human check of the reaction coding is in progress.

The paper counts every panelist who had a political conversation, while the rest of this site covers adults only, so its figures differ slightly from those in Report 2. The panel is not a random sample of the country.

Suggested citation: Ognyanova, K., Aakhus, M., Garimella, K., Jhaver, S., Lee, E. G., Li, J., & Shugars, S. (2026). The elusive verification: How people respond to AI information about politics (Working paper). Rutgers University.

Keep exploring

Report 2Politics on ChatGPT Ask the dataWho asks about politics the most?