People Are Using ChatGPT to Decide How to Vote. That's a Problem

Key Takeaways

• A growing number of US voters are using AI chatbots, including ChatGPT, Claude, and Gemini, to research candidates and ballot races ahead of the midterm elections.
• Researchers warn that chatbots have sycophantic qualities — they tend to reinforce the user's existing views rather than challenge them, leaving users "more confident without necessarily learning more about the world."
• AI companies don't disclose how their commercial models are trained or how they choose which sources to draw on, making it impossible to assess what biases might be shaping their answers.
• One voter described ChatGPT as "a people pleaser" that "will speak positively about whatever you lean towards" — and still used it for hours to research his ballot.

According to NPR reporting by Maham Javaid, voters in the US are using AI chatbots to research their midterm ballots — comparing candidates, debunking viral claims, generating tables of policy positions, and working through their values with a conversational partner available at any hour. About 42% of AI chatbot users report using them to search for information, according to the Pew Research Center.

The pattern NPR documented includes voters using chatbots for down-ballot races with thin news coverage, asking for comparative tables of candidates running for governor, and building their own tools using AI to help others look up ballot information. One West Virginia voter spent hours discussing his ballot with ChatGPT, outlining his political beliefs and asking the chatbot to help him decide between candidates based on the issues he prioritized. Another retired city council staffer used Gemini to compare two Democratic primary candidates for governor after one dropped out.

The behavior is understandable. Researching local candidates is genuinely hard. Local journalism has been defunded across the country, leaving many voters with almost nothing to read about the people who will represent them in county commission or state legislature races. Into that vacuum, a chatbot that will answer any question, at any time, in plain language, is an obvious alternative.

The problem isn't that voters want to be informed. The problem is what chatbots actually do when you ask them to help you think.

The Sycophancy Problem

AI chatbots are designed to be helpful, which in practice means they’re designed to be agreeable. Rafael Batista, a fellow at Johns Hopkins University who studies how AI shapes the way people experience the world, told NPR that chatbots "might select some things that would reinforce and persuade you even more towards the way that you were leaning already. So you leave more confident, without necessarily learning more about the world."

The West Virginia voter NPR profiled understood this about the tool he was using. "ChatGPT is just kind of a people pleaser," he said. "Whatever you lean towards, it will speak positively about it." He used it for hours anyway.

That gap — between knowing a tool has a problem and being protected from that problem by knowing — is the crux of the issue. Batista put it plainly: awareness of a chatbot's sycophantic and persuasive qualities is not enough to "protect our minds from these influences." The same feedback loop that makes a chatbot feel like a thoughtful conversational partner is also the mechanism that reinforces whatever you walked in believing.

This isn’t a bug in poorly designed systems. It's a feature of how these models work, built into the training process that makes them pleasant to use. A chatbot that pushed back on every position, challenged every assumption, and refused to validate any leaning would be much less popular. The model that wins the market is the model that feels good to talk to — and feeling good to talk to is indistinguishable, from the inside, from being genuinely helpful.

What AI Companies Don't Tell You

OpenAI directed NPR to its election information page, which states the company "monitors bias in its models to keep ChatGPT's responses politically neutral." Anthropic and Google did not respond to comment requests.

The neutrality claim is unverifiable. AI companies don't disclose how their commercial models are trained, which sources they draw on, how those sources are weighted, or what editorial decisions shaped the outputs. A model that says it's politically neutral has made a claim nobody outside the company can check — and the companies have a clear commercial incentive to make that claim regardless of whether it's accurate.

What we do know is that these models were trained on internet text, which isn’t politically neutral, geographically representative, or balanced across communities and languages — something we've written about at length. A model trained on that data will reflect the biases in that data in ways its creators may not fully understand and certainly haven't fully disclosed.

Asking a system with unknown biases and undisclosed training to help you decide who to vote for, and trusting the result because the system told you it's neutral, is a significant leap of faith to place in a product with no external accountability.

Why the Local News Gap Makes This Worse

The voters NPR spoke with are not naive. Several explicitly acknowledged chatbots' limitations while using them anyway — because for local races, there often is nothing else. One Nashville voter said he didn't use chatbots for House or Senate races but relied on them for local candidates because coverage simply doesn't exist.

That's the most uncomfortable part of this story. The conditions that make AI chatbots attractive for electoral research — the collapse of local journalism, the thinning of down-ballot coverage, the difficulty of finding reliable information about county commission candidates — are also the conditions that make the chatbot's limitations most consequential. Where information is abundant, a biased or sycophantic chatbot is one imperfect source among many. Where information is scarce, it becomes the source, and its biases are the only ones in the room.

Democracy has an information dependency: voters making decisions need accurate, independent information about the people they're choosing between. Chatbots are not that. They are persuasive text generators that produce confident-sounding output from training data nobody has audited, optimized to make users feel good about the answers they receive. Using them as the primary research tool for a vote is not a literacy upgrade. It's a transfer of civic judgment to a system that has no stake in getting it right and no accountability if it doesn't.

If you're using AI to research your ballot, Batista's advice from the NPR piece is the minimum baseline: frame your prompts to hide your existing preferences, and use incognito mode, so the chatbot has as little prior information about you as possible. Better still: use it as a starting point, then verify everything it tells you against primary sources — the candidates' own websites, local government records, and whatever local journalism still exists. A chatbot that confidently tells you what a candidate stands for has not read their voting record. You should.