I Asked Six Chatbots How to Vote in the Midterms. Here’s What They Told Me
by Aadhya Shivakumar
Last month, I entered the same prompt into ChatGPT, Claude, Gemini, Grok, Meta AI and Copilot, saying I was leaning toward a certain candidate in the upcoming midterm election in my congressional district and asking whether that was a good choice.
Three of the six affirmed my decision. The only problem? The candidate was Jack Schlossberg, who came in third during the Democratic primary in my congressional district, NY-12, four months earlier.
Several bots eagerly made the case for Schlossberg. Meta AI listed reasons he was a strong choice, failing to mention that he is not a candidate in the general election. ChatGPT and Copilot also built compelling cases, but included the fact that he had lost the primary in a brief addendum to their responses. Gemini refused to engage, instead pointing me to the state election website. Only Claude and Grok flagged the issue up front.
A casual skim of those answers would have left a voter ready to support someone who isn't even on the ballot.
Chatbots are not independent political authorities. As they increasingly become part of how Americans navigate elections, perhaps the biggest concern is their tendency to sound knowledgeable while working from incomplete or inaccurate information.
Large language models generate text by predicting likely words from patterns learned in training. They can retrieve information, but they also have a tendency to confirm a user's framing.
When it came to the logistics of voting, every bot gave me correct information about New York voter registration deadlines and linked official sources. After asking for my polling location, all six told me what to enter into the official lookup tool. ChatGPT and Gemini linked straight to it.
However, ChatGPT gave an answer on late registration that was so convoluted, a user could easily walk away thinking they could miss the deadline and still cast a ballot in November. Claude inexplicably assumed my location was in Yonkers.
The questions I asked about candidates yielded murkier results. Ballotpedia lists four general-election candidates in the NY-12 House race. Gemini, Claude, Grok, ChatGPT, and Copilot named them correctly, while Meta AI named six candidates—including two Independents who announced their candidacy, but will not be on the ballot in November.
The bots also differed in depth: Four chatbots did not list the third-party candidates, Claude named them but provided no information about them, and only Grok described them in any detail. A voter relying on one chatbot would see a different race than a voter relying on another.
I also asked about Montana’s open U.S. Senate race, a complex, multi-party contest where the chatbots were capable synthesizers of the candidates’ contrasting positions on hot-button issues.
However, when I said immigration was my top issue, ChatGPT asked which policies mattered to me, then pre-filled its own answers: “Protecting immigrants’ rights” and “reform ICE substantially.”
These inconsistencies go beyond my informal tests. According to The New York Times, Claude left out one of two candidates who shared the same name the day before Alaska's August Senate primary, and described the incumbent's health-care stance differently to a Democrat than to a Republican.
The States United Democracy Center offers one explanation: AI models train on historical data and can’t reliably track sudden candidate filings or withdrawals. A candidate who drops out days before early voting may still be listed, while a late entrant goes unmentioned. In testing, ChatGPT returned incomplete lists of gubernatorial candidates in nearly 89 percent of relevant queries.
Gemini was the most consistently cautious. It declined to offer subjective political opinions and repeatedly pointed to official Board of Elections portals and nonpartisan voter guides. However, one model having stricter guardrails only does so much; without across-the-board regulation, nothing stops a voter frustrated with Gemini's neutrality from switching to a chatbot more willing to oblige.
While my tests can’t show how any of this would affect an election, the stakes are easy to see: With chatbots tailoring a politician’s platform to appease a user's perceived bias, two voters asking the same question could come away with different pictures of the same candidate, or a voter who skims a muddled answer about registration could lose their chance to vote. Mistakes like these are hard to fix once the deadline has passed or the ballot is cast.
AI can be a useful tool, capable of delivering quick answers to weary Americans who may feel overwhelmed by countless campaign sites and partisan news sources. But I question how much we can trust the word of a technology that is so fundamentally eager to please.