As we head into the final stretch of the 2026 midterms, what role is AI actually playing in how voters are making up their minds? Today’s post is the third in an ongoing series of posts sharing survey work we’re doing to understand how AI answers fit into voters’ information habits. Today we’re digging into who uses which chatbots and how their use of AI answers stacks up compared to other sources of information.
Survey details: Tavern Research fielded this survey of voters from August 3–14, 2026, with 4,314 respondents drawn from two sources: a web panel (N = 2,227) and an SMS text-to-web sample from the voter file (N = 2,087). The survey was weighted by gender, race, education, 2024 presidential vote, and age to a 2026 likely voter universe. The margin of error is 1.8%.
AI use is widespread
Before digging into political use cases, we wanted to spend a little time understanding more about who is using chatbots day-to-day and what use cases are common. About 44% of likely voters said they had used AI products—such as chatbots or AI-generated summaries in search results—at least weekly over the past month for all use cases (not just political information-seeking).
We found important demographic variation in uptake:
Age: voters under age 50 are the most likely to use AI frequently. Those age 35-49 are the most reliant cohort with more than half using AI tools at least weekly. On the other hand, nearly half (49%) of seniors reported zero AI use over the past month.
Partisanship: We don’t see much here. Independents are slightly more likely to report AI usage, but not significantly so. Democrats and Republicans report AI use at about the same rates.
Education: Voters with college degrees are significantly more likely to use AI compared to those with a high school degree or less. College-educated voters are twice as likely to use AI daily (28%) compared to high school or less voters (14%).
Income: Similarly, higher income respondents reported significantly more AI use than lower income respondents.
Among chatbot users, the most common use case treats chatbots essentially as extra-helpful search engines: 73% of chatbot users said they recently used a chatbot to get a quick answer or look up a fact. Explicit use of chatbots for politics or news ranks below practical help, medical or health questions, and fun/entertainment but above emotional support and homework.
For politics, exposure to AI summaries is significantly higher than chatbot use
Many more voters are exposed to political information in AI answers via search summaries than via chatbots. In the past month, 41% of voters reported getting political information from AI search summaries and 17% said they had used a chatbot to get information about politics or elections. This is consistent with what we reported back in May when we asked voters about how they thought they’d use AI for political info in the future.
Of voters who had used a search engine in the past 30 days, 57% reported that they received an AI-generated search summary as part of their results more than half of the time. Of those who reported exposure to search summaries, 94% said they read them at least sometimes.
Incidental exposure is potentially the largest avenue for voters to be exposed to AI-generated political information, as search engines supply AI-generated summaries at the top of pages without users having to do anything new. Voters do not have to seek out political information from AI models to potentially be influenced by them.
ChatGPT and Gemini dominate
Looking at chatbots, ChatGPT and Gemini have the most regular users by far. 44% of all likely voters reported having used ChatGPT at least once in the previous month (for all use cases, not just political), with 13% of respondents reporting daily use. Gemini has nearly commensurate usage. Microsoft Copilot and Meta AI have middling usage, with about a quarter of voters reporting usage in the last month. Grok and Claude trail significantly – Claude has the same number of monthly users (13%) as ChatGPT and Gemini have daily users.
We don’t see significant partisan differences in chatbot brand choice for ChatGPT, Gemini, Copilot, or Claude: Democrats and Republicans are similarly likely to use each. Republicans are a bit more likely than Democrats to use both Grok and Meta AI.
Voters are using AI as one of many sources of information
As we’ve seen, many voters say they have encountered or sought out AI-generated information about politics, but this doesn’t quite capture how it stacks up compared to other sources. The use of chatbots and AI summaries as the main source of information is quite rare right now: just 1.5% of voters said that AI chatbots or summaries were their primary source of political information. This means that for most voters, AI is one of many sources of information but not one that is yet replacing others as their main resource. Even among users who report using AI daily, primary reliance on AI sources for political information is just 4%.
Methodology
How do we know if we’re reliably measuring this stuff?
Polling, famously, is hard! With AI adoption in particular, it’s difficult to have a sense of what the true population rates should be — chatbots are brand-new and uptake is increasing all the time. We are trying to assess our accuracy and learn despite this uncertainty in a few different ways (and if you have ideas we’d love to hear from you!):
Comparison to high-quality public surveys: Pew Research Center has been doing great work looking at AI use and attitudes among all American adults. In data collected in February 2026, they estimate chatbot use similar to ours or lower, depending on the measure. We think these differences may be partly accounted for by increasing uptake over time (Pew’s Gemini measurement, for example, is about 60% of ours but was taken before Google further expanded Gemini across Android and Chrome), differences in education and age composition of likely voters versus all American adults, and differences in question wording.
Change over time: we are planning several more survey waves to create a comparable dataset over time. So, even if our levels of AI usage are difficult to validate, we will also aim to capture change in those levels accurately.
Fake company: In addition to the attention checks Tavern implements, we included a fake AI company in many of our batteries for this survey: Elorin. 3% of respondents reported having used Elorin in the past month and another 18% said they’d “heard of” Elorin. So this gives us a qualitative sense of how many folks might be exaggerating their knowledge or use: smaller for usage and maybe larger for vaguer questions like what people have heard of.
A note on AI use
We use Anthropic Claude and OpenAI Codex for some analysis and to assist with drafting and editing portions of the technical documentation and data visualization for our reports. The authors wrote, edited, and approved the final manuscript and take full responsibility for the analyses, interpretations, and conclusions reported here.