The NYT is right: voters are asking AI who to vote for. Here's what the chatbots are actually telling them.

The New York Times reported this weekend on voters who are taking photos of their ballots and asking Claude and ChatGPT for information on the candidates, strategic advice, and even recommendations of who to vote for. They note that “The 2026 midterms may be the first American elections in which voters are using A.I. in meaningful numbers.”

We’ve been measuring voter AI use, as well as what chatbots are telling voters, and we agree that AI chatbots will be a real factor this fall – for the first time in an American general election. The NYT’s interviews are data points in a larger overall pattern we’ve built Caucus AI to uncover. In today’s post, we review what we’ve learned so far and how it dovetails with what the Times found.

According to a survey we conducted with Change Research in May, a significant minority (15%) of voters say they are very or somewhat likely to consult an AI chatbot for information on midterm candidates. A significantly larger share, 62%, say they are likely to use Google Search – and so may be exposed to AI-generated political information via Google’s AI search overviews.

Two of the most persuadable groups of voters in recent elections are more likely to turn to chatbots. 21% of voters under age 35 say they may use chatbots, compared to just 8% of voters over 65. Voters of color are nearly twice as likely to say they’ll use AI chatbots (21%), compared to 12% of white voters. This represents a small share of the electorate, but often a decisive one.

It isn’t just persuadable top-of-ticket voters who are turning to chatbots though – more engaged voters participating in primaries or seeking information on less prominent, downballot races are also consulting these tools, as the NY Times article showed in the California primaries. In our survey, respondents who were decided on the generic Congressional ballot were actually more likely to say they would use a chatbot (15% of Democrats and 17% of Republicans), compared to just 9% of undecided voters. This is consistent with the fact that stronger partisans are more likely to seek information across all channels and in 2026, that includes AI chatbots.

The voters the Times profiled also asked questions that track with our national data – strategic, subjective, and often personal questions. In our Change Research survey we asked respondents for the actual questions they’d ask a chatbot. Of substantive responses, 39% asked questions requiring chatbots to make judgment calls or synthesize across multiple sources (e.g. “Which candidate is more likely to tell the truth and practice what they preach?”). More than half of responses from voters under 35 were these types of questions. 18% of substantive responses overall were asks for the chatbot to compare and contrast candidates or recommend who to vote for.

Since February, we’ve been monitoring what information voters get back from chatbots when they ask about politics. We’ve been querying ChatGPT, Gemini, and Grok with identical prompts about 2026 midterm candidates at the House, Senate, and gubernatorial levels. This allows us a bird’s eye view of what chatbots are saying about the 2026 midterms. Findings so far include –

• Neutrality is the goal: As the New York Times also found, frontier labs creating these chatbots are striving for neutrality – in our data so far, we have found that models are clearly orienting around this goal. Models present generally positive views on bio and issues, unless asked specifically about criticisms of a candidate.

• Sourcing: Our core finding is that the sources chatbots draw on vary widely from model to model, and even within model runs. Models appear to heavily favor certain publishers–many of which have content partnerships with the labs–and can ignore otherwise prominent outlets and sites. This means that when a user selects a model, they are unknowingly selecting into a specific information ecosystem.

• Answers update quickly: Because most frontier AI chatbots use web search to respond to questions about elections, it means they have the ability to incorporate new content as soon as it goes live online. This is helpful with breaking news – while monitoring California’s governor’s race, we found that chatbots picked up information about sexual misconduct allegations against Eric Swalwell the same day news stories broke. On the other hand, less well-vetted user-generated content may also be picked up quickly – in March, we created two Wikipedia pages for candidates we thought met Wikipedia’s notability regulations, but didn’t yet have pages. This content was picked up by ChatGPT in under 15 minutes.

• Incumbency and prominence matter: As Yamil Velez pointed out in the article, chatbots favor candidates with bigger media footprints. We quantified this with a similarity exercise: comparing chatbot responses to candidate websites (using the website as a proxy for the message the candidate is trying to deliver to voters). We found that chatbot responses were more similar to candidates’ websites for incumbents and more prominent candidates. Challengers and less prominent candidates have a harder time getting their message across via chatbot.

All of this is dynamic – frontier AI labs are developing policy and partnerships in real time, chatbots increasingly rely on memory of users’ preferences, chatbot use and voter attitudes toward AI in general are rapidly changing, and campaigns themselves are exploring ways to adapt to an increasingly AI-generated information ecosystem.

We will be expanding our work with Caucus AI to explore these open questions and monitor change over time as November approaches. If you have ideas for what we should look into next or want to dig into what chatbots are saying about a specific race, please get in touch!