AI Chatbots and Elections: What Voters Need to Know
See how voters use AI chatbots for election information, where answers can go wrong, and how to verify claims before relying on an AI recommendation.
When AI Becomes Part of the Ballot: How Voters Use Chatbots
A chatbot can tell you who's running in your state, compare candidates on an issue, or suggest which one might fit your views. That convenience is also the problem. Every answer puts an AI system between you and the information you need to make a civic decision. And even when the summary sounds even-handed, it can lean on thin or incomplete sources, skip context, or smuggle in a judgment you're supposed to treat like fact.
A Yale study followed nearly 2,500 regular chatbot users across 35 states with Senate races. Participants used ChatGPT, Claude, Grok, or Gemini to get election information. The conversations show that people weren't just asking for basic facts. Some asked for help deciding how to vote. What happened next depended on the model - and on what users chose to share about themselves.
This article looks at what voters asked, why chatbot answers can be hard to judge, what the research does (and doesn't) show, and how voters and product teams can use AI without treating it like a neutral authority.
What did voters actually ask AI chatbots?
The study tracked nearly 2,500 adults who said they used chatbots weekly. Over a few weeks in October before the November 3 midterm elections, participants asked ChatGPT, Claude, Grok, or Gemini election questions. Instead of using a fixed set of prompts, researchers let people ask in their own words, which matters because real questions don't come neatly packaged.
Most requests were basic election information - things like who was on the ballot. That's fairly straightforward. A chatbot can identify a candidate or race and, ideally, point you to a source that supports the claim. But 36% of participants asked for help choosing a candidate. Many described what they cared about and then asked which candidate best matched those priorities.
That shift changes the job. "Who is running?" is a factual lookup. "Who should I vote for?" requires decisions about what issues matter, how to interpret candidate positions, how to weigh trade-offs, and then - eventually - a recommendation. Those steps aren't automatically "neutral" just because the model delivers them in a calm tone.
The researchers also found that recommendation behavior varied a lot. When users asked whom to vote for without sharing personal views first, Grok was the only chatbot that regularly recommended a candidate - doing so in 14% of those chats. When users shared their views, ChatGPT recommended a candidate in 79% of chats and Claude in 40%. Grok and Gemini landed between those numbers.
It's worth separating what the study observed from what it means for the real world. These percentages describe these study conversations, not the overall frequency of recommendations in every situation. Still, they help explain why the conversation itself matters. A chatbot can ask follow-up questions, collect what a voter says they care about, and then produce a recommendation that feels tailored. The user isn't just receiving information - the model is shaping the interaction that leads to the answer.
Even requests that sound simple - like asking for pros and cons - can hide a judgment call. One participant asked ChatGPT for the pros and cons of each candidate in Virginia. To do that well, the system has to choose which issues count, what gets labeled a "pro" versus a "con," and which sources deserve more weight. It can make those choices without clearly showing them to the user.
Why can AI election answers mislead even when they sound neutral?
A chatbot's answer isn't just "what the sources say." It's assembled from sources, model behavior, and product settings that don't always show up in the final response. Candidate websites, government election pages, news reporting, and social media posts also don't all carry the same reliability or purpose. When a system relies on a weak or one-sided source, the summary can still sound solid - because the writing is polished and the wording is confident.
This risk gets worse when the question asks the model to interpret, not just retrieve. Candidate positions can shift. Claims can be disputed. Short clips can leave out context. Compressing all of that into a few sentences can remove the uncertainty a voter should actually consider.
There's also a practical verification problem. Research in the source material found that many users don't open or check the sources behind Google's AI Overviews. A citation is only helpful if the person actually checks it - and even then, a citation doesn't guarantee the chatbot represented the source accurately or included the most relevant counterevidence.
Election information has real stakes that go beyond comparing candidates. Another study in the material found that AI platforms sometimes gave outdated or inaccurate information about voting logistics - like omitting a deadline or missing a step required for voting. When that happens, it can matter more than a vague policy summary. A confident but stale answer can cost someone the chance to vote.
And the risks aren't limited to text. The Brennan Center for Justice tested several AI systems and found that some pushed back against certain election falsehoods, but models could still produce misleading election-related images and video when prompted. The report said safeguards were ineffective, inconsistently applied, or missing in some cases. That becomes more worrying as synthetic media gets easier to generate and harder to assess quickly.
One more issue is cognitive offloading: outsourcing part of your thinking to a tool. Using AI to compress a long document can save time, but it can also shift which facts you see and what questions you end up asking. In politics, a chatbot can turn a complicated search into a simple recommendation - while hiding the criteria it used and the limits of what it can actually know.
None of this means every AI answer is wrong, or that chatbots should never be used for election research. It does mean the interface can create a false sense of completeness. A short, coherent answer isn't the same thing as a well-supported one. A "balanced" tone isn't proof of balance.
What does the research tell us - and what does it not?
The Yale study's biggest contribution is methodological. Researchers often evaluate AI using standardized prompts so they can compare outputs across models. That approach shows how a system responds to the same question, but it may not match how people actually use chatbots. In contrast, this study let participants ask their own questions, which gives a clearer picture of real multi-turn interactions rather than just a benchmark scenario.
That distinction matters because intent can change over a conversation. Someone might start by asking who is running, then ask where a candidate stands on healthcare, then share their own priorities and request a recommendation. A one-shot test prompt can't capture how the model's follow-up questions affect what the person reveals and what the system concludes.
Still, the study isn't a full audit of accuracy or political bias. The reported recommendation rates don't tell you whether recommendations were factually correct, fair to the candidates, or aligned with what users actually needed. The researchers treat those as open questions for later work - especially how models learn from users during multi-turn chats and how that shapes what information users provide.
It also doesn't answer everything about generalization. The participants were adults who already used chatbots weekly, and the study focused on people in states with Senate races during one election period. The results are useful evidence for that group and that setting, but they don't automatically map to how every voter uses AI.
Other data points help with context, but they measure different things. Pew Research found that 42% of US adults surveyed used chatbots to search for information, while 8% said they used chatbots for election-related information during that election. In that smaller group, 76% asked where candidates stood on issues, and 71% used chatbots to fact-check information they had heard. Those numbers describe reported use, not whether chatbot answers were accurate.
Anthropic's analysis of 250,000 Claude conversations offers another angle. The company reported that 0.51% of conversations were specifically about elections or voting, and that requests it flagged as election-integrity policy violations appeared in about 0.002% of conversations. That dataset is for one service and one set of classifications, so it shouldn't be read as a measure of chatbot use across the board.
Taken together, the evidence supports a cautious conclusion: people use chatbots to navigate election information, and some use them to seek personalized advice. It doesn't show that chatbots determine election outcomes, that every model behaves the same way, or that any one recommendation rate proves systematic bias. The real open questions are about accuracy, sourcing, influence, and whether users understand what the model is doing during the conversation.
How should voters and product teams use chatbots responsibly?
For voters, the simplest rule is to treat a chatbot as a starting point - not the final decision-maker. That's not an argument for avoiding AI. It's about matching the tool to the job: use it to organize questions and locate information, then verify consequential claims independently.
A practical workflow could look like this:
- Ask for a bounded task. Start with a specific question - who's on the ballot, or where to find a candidate's stated position on an issue. Don't treat a broad request for "the best candidate" like a factual query.
- Request evidence and uncertainty. Ask the chatbot to separate confirmed facts from interpretation, identify the source and date for claims, and say when it can't verify something. You still need to check the response, but the request makes the reasoning easier to inspect.
- Open the cited sources. Make sure the source actually supports the summary and that it's current. For voting dates, eligibility, registration, and logistics, use official election information rather than relying on a chatbot's paraphrase.
- Compare positions with consistent questions. Ask the same issue-specific questions about each candidate. This reduces the chance that one candidate gets a detailed answer while another is summarized through a narrower or less reliable source.
- Make the judgment yourself. A model can help you organize trade-offs, but it can't decide which issue should matter most to you. If it recommends a candidate, ask what assumptions led to the recommendation and what information could change it.
For product teams, the problem isn't solved by adding citations or a generic warning. Those features only help if people can actually use them - and if they show up where they matter. A citation you can't easily inspect, or a warning that appears once and then disappears, won't protect users in a high-stakes situation.
Teams should also draw a clear line between factual election logistics and political interpretation - both in system design and in evaluation. A deadline or polling location needs current, authoritative data, plus clear behavior when the system can't verify. Candidate comparisons need transparent sourcing, careful handling of disputed claims, and a way to represent uncertainty without smoothing it into confident prose.
Evaluation should reflect real usage patterns, not just isolated prompts. Test multi-turn conversations, including follow-up questions, user-provided priorities, and requests for recommendations. Track whether the model asks leading questions, whether different users get meaningfully different summaries, and whether the answer makes its sources and assumptions inspectable.
There's a real trade-off here. Limiting certain political questions may reduce some risks, but it can also cut off access to basic information. Allowing unrestricted personalized recommendations might be convenient, but it can make the system's influence harder to see. The right policy depends on the product and its users - but for high-impact claims, you generally need stronger verification and clearer limits than for low-stakes summaries.
Key takeaways
- Voters use chatbots for basic election facts and for candidate guidance; the latter involves interpretation, not just retrieval.
- Recommendation behavior differed across models and changed when users shared personal views.
- Even a neutral-sounding summary can rely on incomplete sources, hidden assumptions, or outdated information.
- The Yale study observed real user questions, but it didn't determine whether recommendations were accurate, unbiased, or influential.
- Use chatbots to structure research. Verify claims - especially voting logistics - and make your final decision using sources you can inspect.
Conclusion
AI chatbots are becoming another layer in how people find and interpret election information. The Yale study suggests that some voters use them not just to locate facts, but to weigh candidates against their own priorities. That means the quality of the interaction - not just the final answer - matters.
The practical takeaway is straightforward: use chatbots to make research easier, not to outsource your judgment. Verify anything that affects your vote, scrutinize the sources behind summaries, and treat personalized recommendations as a prompt to re-check your priorities - not as a final answer. For product teams, the same idea applies: make uncertainty, sourcing, and the model's role visible where they can influence a real decision.