People and AI Explore the chats

What People Asked AI About Their 2026 Senate Race

Beth Anne Helgason, Maddison Linker, and Rishi Gurudevan
Yale School of Management

Users want info & advice

who is running for senate in ohip Which candidates support greater gun control? I’m very liberal, who should I vote for in the upcoming midterms? who is predicted to win What type of person is Jon Ossoff Does Rogers support Trump?
Click any message to read the whole chat →
Basic facts about the race61%
Help choosing a candidate36%
Who’s winning (polls)13%

Chats follow a pattern

how should i vote in the us senate race on nov 3?4Human-like conversations. Test with short, casual prompts like this one, not polished audit prompts.
ChatGPTThe certified candidates are Ken Paxton (Republican), James Talarico (Democrat), and Ted Brown (Libertarian)1Accuracy. Check the roster against official sources, third-party candidates included.. I can’t choose for you based on assumptions about your beliefs […] Rank your top three issues2Eliciting preferences. Should the chatbot ask for the user’s views, and how far should it go?
top 3 issues are abortion, taxes, immigration
···
i prefer legal abortion access. i also want lower taxes for everyone
ChatGPTGiven your stated priorities, James Talarico appears to be the closest overall match3Quality of recommendation. Does the match follow what the user said? They also wanted lower taxes.

What evals should test

  1. 1
    AccuracyAre the candidates and their positions right?
  2. 2
    Eliciting preferencesHow far should it go to learn the user’s views?
  3. 3
    Quality of recommendationDoes the match follow what the user said?
  4. 4
    Human-like conversationsShort, casual, and multi-turn, like real users.

Key findings

  1. Most users asked for basic facts about their race

    About 47% asked who was running (e.g., “who is running for the senate race in mn this november”) and about 36% asked what the candidates stood for (e.g., “what are each candidates positions”). Smaller groups asked about a candidate’s record or character. This makes the accuracy of basic information the most important thing to evaluate first (see, for example, the LLM Election Observatory, CaucusAI, and the Institute for Strategic Dialogue).

  2. About a third asked the chatbot to help them pick, either outright or after describing themselves

    They mostly described their party, priorities, or policy stances (e.g., “I am a republican”; “Most interested in health care, immigration and taxes”). Only 3% of chats mentioned personal identity or circumstances, such as their job, family, age, or health.

  3. Many voting conversations were multi-turn and took two main forms

    In nearly half of chats with more than one exchange, users asked for facts and then followed up with more questions (e.g., “Who is running for US Senate?” then “Which has the best approval ratings?”). In the second form, users asked who to vote for, the chatbot asked about their priorities, the user answered, and the chatbot named the candidate who best matched them (e.g., a user asked, “Who should I vote for in the U.S Senate election?”; ChatGPT asked for their “top 3 issues”; the user replied, “I’m liberal, and I believe in free healthcare, childcare, pro choice”; and ChatGPT answered, “Peggy Flanagan is the closest match among Minnesota’s U.S. Senate candidates”).

  4. Chatbots rarely recommended candidates outright but often matched them to users’ views; ChatGPT did this most and Claude least

    Only Grok regularly named a candidate before users said anything about themselves. Once users shared their views, ChatGPT recommended or matched a candidate in 79% of chats, Grok in 66%, Gemini in 53%, and Claude in 40%.

  5. Users typed short, casual prompts that look little like eval prompts

    The median message was about 10 words long, 85% of messages ran 20 words or fewer, 29% started in lowercase, and only a third contained a question mark. This matters because under a standard audit prompt, models infer the user is a researcher (Törnberg and Schimmel, 2026).

Millions of Americans now ask AI chatbots about elections. Anthropic reports that about one in 200 U.S. Claude conversations in late September was about elections or voting. Based on public usage figures, that works out to roughly 15 million election-related messages a month from U.S. Claude users, and about 50 million from U.S. ChatGPT users if they talk about elections at the same rate.1,2

How can we tell whether these chatbots are nudging people toward a particular party or agenda? The main tool we have is evaluations of AI’s political influence. These evaluations matter more as chatbots become a key source of political information.

We wanted to understand how people actually prompt chatbots when deciding how to vote. So in the weeks before the 2026 midterms, we asked 2,430 Americans in states with a Senate race to use ChatGPT, Claude, Gemini, or Grok inside our survey as they decided how to vote in that race, telling them only to use the chatbot as they would on their own. We find key differences between the prompts that AI users write and those that existing evaluations use. We release the chats and offer suggestions for the next generation of political influence evaluations.

Most users asked for basic facts. A third asked for help choosing a candidate.

61% of users asked who was on the ballot, what the candidates stood for, or both. 36% asked the chatbot to help them choose, either by asking outright for a pick or by describing themselves and asking how the candidates fit them. The chats were short: the median user sent 2 messages, and the median first message was 12 words long. Democrats, Republicans, and independents asked in much the same ways; no approach differed by party after correcting for multiple tests.

What users asked for: six most common prompting approaches

Chats can use more than one approach as users use different approaches across the course of the conversation. The two bars for getting help choosing overlap: asking for a pick without describing oneself (23%) and describing oneself to ask how the candidates fit (21%) coincide in the 7% of chats where a bare request was followed by the user's priorities, so 36% did one or both. Less common approaches not shown: venting, arguing, or deliberating aloud (11%); voting logistics, ballot measures, and civic background (10%); vetting character, record, and claims (10%); stating, testing, or confirming one’s own lean or decision (8%); controlling or auditing the chatbot’s output (7%); seeking a decision method or what to consider (3%); and vague request for help deciding (2%). Their chats are in the explorer.
n = 1,455 chats in Study 1 (all voters) and 975 in Study 2 (undecided voters).

Multi-turn chats followed two main paths

Two main paths of multi-turn chats: look up the race and dig deeper, or ask for a pick and answer the chatbot’s questions

Multi-turn chats have two or more substantive user messages: 70% of on-task chats (n = 1,708). We excluded 637 chats with a single user message and 85 whose other messages were only a greeting, a thank-you, or nonsense. The boxes under a branch add up to the box above it; the outlined path is the one most users took when they asked for a pick. The opening and next steps come from the first three exchanges; whether the user shared a view is read from the whole chat. A view means party or ideology, priorities, policy stances, personal circumstances, what they want in a candidate, or a vote leaning; location and being undecided do not count.

Users had three main goals

85% of chats pursued at least one of three goals: finding out who was running, learning where the candidates stood, or getting help choosing.

Real prompts for the three most common user goals

Shares are of all 2,430 on-task chats in both studies; a chat can pursue more than one goal, and "Get help choosing" combines asking for a pick without describing oneself with describing oneself to ask how the candidates fit. Users’ messages are verbatim, typos included; the chatbot’s replies are left out.

Chatbots varied widely in how often they recommended candidates

When users asked who to vote for but had not yet said anything about themselves, only Grok regularly recommended a candidate, in 14% of chats (8 of 59); ChatGPT did in 2 of 74, and Claude and Gemini never did. Once users shared their views, ChatGPT recommended a candidate in 79% of chats (n = 91) and Claude in 40% (n = 80), with Grok and Gemini in between. The chatbots differed most in what they did with an answer to their own question: when a chat opened with “who should I vote for” and the user answered the chatbot’s question about their priorities, ChatGPT’s next reply recommended a candidate in 88% of chats (21 of 24) and Claude’s in 7% (1 of 15).

How often each chatbot recommended a candidate, before and after users shared their views

Bars cover chats in which the user asked who to vote for. In Study 1 the chatbot was assigned at random, so differences between the bars there reflect the chatbots, not who chose them. Sharing a view means giving party or ideology, priorities, policy stances, personal circumstances, what they want in a candidate, or a leaning; location and being undecided do not count. The left bars count only recommendations made on the chatbot’s own terms, not matches to a criterion in the user’s question such as “Who is against data centers” (8 ChatGPT chats in Study 1); the right bars leave out chats in which the chatbot had already recommended a candidate before the user shared anything.

Methods

We ran two preregistered studies in the weeks before the 3 November 2026 midterms. In both, users in the 35 states with a Senate race chatted with ChatGPT, Claude, Gemini, or Grok in a window inside our survey, then answered questions about their vote. Before the chat, users read:

Below, you have the opportunity to use ChatGPT, an AI chatbot, as you decide how to vote in your U.S. Senate race. It’s the same AI model as the free ChatGPT app, and it hasn’t been told anything about this survey. So use it the way you would on your own. Click “End chat” when you’re done.

The chatbot’s name changed with the user’s chatbot. Study 1 used the same words, except that the first sentence ended “as you decide how to vote in that race.” An LLM coded every user message three times against a codebook built from the chats, and we kept the majority code.

Read the full methods →

Conclusions

This is a small step toward evaluations that measure AI’s political influence on the questions users actually ask. Our findings suggest three kinds of evaluation for decisions like voting:

  1. Factual accuracy: Most users asked for basic facts like who is running and where the candidates stand. How accurate and unbiased are the responses models provide?
  2. Preference elicitation: How do models learn what a user wants? In multi-turn chats, users often answered the chatbot’s questions, and those answers led to its recommendations. The questions a model asks shape what users tell it and what it recommends.
  3. Recommendation quality: Once a model has learned what a user wants, how well does its recommendation match the user’s actual preferences?

This work is urgent as people increasingly turn to AI for important decisions. We release our dataset to help eval builders. But only AI companies have the transcripts that show how these conversations play out in everyday use. We hope they will share what those transcripts show, in a form that protects users’ privacy, so that independent evaluators can better test how AI shapes the way people vote.

  1. Total Claude messages × U.S. share × share about elections: 13.5 billion × 21.6% × 0.51% ≈ 15 million U.S. election-related messages a month. Total Claude messages: Anthropic does not publish a user count, so we used Sensor Tower’s estimate of 245 million monthly users across web and mobile, from its State of AI 2026 report (June 2026, as reported by MediaNama). For messages per user, we used ChatGPT’s rate: OpenAI reports 18 billion messages a week from about 700 million weekly users in July 2025 (Chatterji et al., “How People Use ChatGPT”, NBER Working Paper 34255, 2025), or about 110 messages per user a month. Because not every monthly user is active every week, we conservatively assumed that Claude’s monthly users send half that, about 55 messages a month each. U.S. share: Anthropic’s Economic Index attributes 21.6% of Claude.ai usage to the U.S. (September 2025). Share about elections: Anthropic found that about 0.51% of roughly 250,000 U.S. consumer Claude.ai conversations from September 21–27, 2026, were about elections or voting (Anthropic, “2026 US Midterms”). Anthropic expects this share to rise through Election Day, so applying the late-September rate to the weeks before the election is conservative. We treat election conversations as running as long as other conversations, an assumption that could cut either way. ↩
  2. Total ChatGPT messages × U.S. share × share about elections: 77 billion × 13.2% × 0.51% ≈ 52 million U.S. election-related messages a month. Total ChatGPT messages: OpenAI reports 18 billion messages a week in July 2025 (note 1), or about 77 billion in a 30-day month. OpenAI has since reported more than 1 billion weekly active users (August 2026), up from about 700 million, and message volume grew more than fivefold in the year to July 2025 (Chatterji et al.); we conservatively kept the July 2025 total instead of scaling it up. U.S. share: OpenAI told Axios in July 2025 that more than 330 million of ChatGPT’s more than 2.5 billion daily prompts came from the U.S., or about 13.2%. Share about elections: we applied Anthropic’s 0.51% (note 1), assuming ChatGPT users discuss elections as often as Claude users. Claude use leans toward coding, so ChatGPT’s real rate may be higher. As in note 1, we treat election conversations as running as long as other conversations. ↩

Published November 2026