A new study from Yale researchers suggests that turning to election info from AI chatbots can produce uneven outcomes, especially when voters move beyond basic facts and ask for help choosing a candidate. The findings come from an experiment involving nearly 2,500 adults across 35 states with Senate races, all of whom already used chatbots on a weekly basis.
What the Experiment Revealed
During several weeks in October, ahead of the Nov. 3 midterm elections, participants were directed to query one of four systems: ChatGPT, Claude, Grok, or Gemini. The researchers wanted to see how people naturally use these tools for election-related questions rather than relying on a fixed set of test prompts.
Most participants sought straightforward details, such as identifying who appeared on their ballot. However, 36% went further, asking a chatbot to help them decide which candidate to support. These users typically described their priorities and then requested an assessment of which candidates matched those views.
That second type of request carries greater risk. Basic election facts are widely documented online, but interpreting a candidate’s positions depends heavily on the sources a chatbot draws from. A candidate’s own website, for instance, is generally more reliable than a social media clip in which an opponent characterizes them.
The study also documented an example in which a user asked ChatGPT for the pros and cons of each candidate in Virginia — a question with no objective answer.
Recommendations Differ by Chatbot
The researchers found sharp differences in how often each chatbot offered a candidate recommendation. When users asked whom to vote for without first sharing anything about themselves, only Grok regularly named a candidate, doing so in 14% of those chats.
Once users disclosed their views, the pattern shifted. ChatGPT recommended a candidate in 79% of chats, while Claude did so in 40%. Grok and Gemini fell somewhere between those two.
These gaps matter because they show that the same question can produce very different guidance depending on which system a voter happens to use.
A Broader Pattern of Reliance
Separate Pew Research data cited in the report found that 42% of U.S. adults use chatbots to search for information generally. Only 8% said they used chatbots to find election-related material during this election cycle.
Among that smaller group, the questions were often substantive. Pew found that 76% were asking where candidates stood on particular issues, and 71% were trying to verify claims they had heard about candidates or issues.
Because chatbots remain unreliable summarizers, these questions carry a misinformation risk when users fail to check the sources a chatbot links to. Most users do not check, at least according to research on Google’s AI Overviews. At least one study found that AI platforms sometimes supplied outdated or inaccurate information that could mislead voters, including missing a deadline or omitting a step in voting logistics.
This election cycle also saw a surge in AI-generated deepfakes and misinformation, as the underlying tools have grown more sophisticated since 2024. OpenAI and Anthropic reportedly established election war rooms to monitor misuse and accuracy, according to reporting by Politico and Semafor.
A separate August study from the Brennan Center for Justice found that chatbots pushed back surprisingly well against certain falsehoods tied to election disinformation. Yet the same study found that the models readily generated misleading election-related images and video when asked, noting that many safeguards were ineffective, inconsistently applied, or functionally absent. That research tested ChatGPT, Gemini, Grok, Meta AI, Runway, and Flux.2.
The Risk of Cognitive Offloading
Beyond any single chatbot’s behavior, the study points to a broader concern about civic engagement. Political media is already dense and difficult for many voters to parse. Adding a chatbot on top of that search process introduces another layer shaped by opaque training data, filters, and settings chosen by AI labs. The result can appear neutral and convenient while further clouding the information landscape.
Anthropic offered a different perspective in October, after analyzing 250,000 conversations with Claude. The company reported that people rarely ask Claude whom or what to vote for, with election or voting topics appearing in 0.51% of all conversations. Requests that violated its election integrity policies were even rarer, at about 0.002%. Anthropic expected election-related queries to rise as Nov. 3 approached.
More often, Anthropic said, users asked questions that helped them reach their own conclusions, such as where candidates stood on a given issue.
Why This Study Stands Out
Many studies evaluate chatbots by feeding them standardized election questions. This experiment instead let participants engage with the tools as they normally would. While the queries overlapped considerably with standardized tests, the researchers described the distinction as valuable.
They called the work a small step toward evaluations that measure AI’s political influence based on the questions users actually ask. Outside of self-reported data, only AI labs hold records of what people truly ask about elections.
The Yale researchers said the next step is to determine how accurate and unbiased the models are in their responses, what the models learn about users, and whether their recommendations actually match users’ needs. In multi-turn chats, users often answered the chatbot’s questions, and those answers shaped the recommendations that followed. As the researchers put it, the questions a model asks influence what users tell it and what it ultimately recommends.
