MIT Researchers Are Asking AI Chatbots the Same Election Questions Every Few Days. The Answers Change Depending on Who Is Asking

With the midterm elections three days away, a team of MIT researchers is running the same set of election questions through popular AI chatbots every few days and publishing what comes back, to document an information environment that is otherwise invisible to anyone outside a single conversation, according to an interview with the researchers published Friday by MIT News.

The project, called the LLM Election Observatory, is a public dashboard led by political science professor Adam Berinsky, Sloan School assistant professor Chara Podimata and political science professor Charles Stewart III. It asks a fixed set of questions on topics including election certification, hot-button issues, result predictions and news coverage, and varies the simulated user's location, political party, identity and other factors to see how responses differ.

"People are increasingly using AI chatbots to find and make sense of political and election-related information," Podimata told MIT News. "That makes it important to understand what information these systems actually provide, not just at one moment in time but throughout an election cycle."

Part of the problem is that there is no historical record. "If an AI system gives a particular answer about an election today and gives a different answer several weeks later, it can be very difficult to reconstruct after the fact what voters would have seen at each point in time," Podimata said. One goal of the project is simply to document that environment and make the data available to other researchers and the public.

Accuracy is only part of it. "Factual accuracy is only the first frontier," Podimata said. "Differences in framing, emphasis, uncertainty, and cited sources can matter enormously, even in the absence of outright misinformation."

The researchers have not finished a full analysis, but Podimata described early patterns in the 2026 data. "Across the different user identities we present, we see differences in how the answers are presented, in which sources a model cites, and in tone — sometimes even when the underlying facts are the same," she said. "What we cannot yet say is how stable those differences are, how they move over the course of the election, or how a real person would read and react to them."

Podimata compared the situation to social media, where researchers could at least partly see what circulated in different groups and communities. "With chatbots, these information environments are largely invisible to anyone outside the individual conversation," she said. "Two users may walk away with very different perceptions of the same underlying facts, without ever realizing that their experiences differed."

The project builds on Podimata's research during the 2024 presidential election, which found meaningful variation in AI models' responses over time and across how prompts were framed, as well as differences in how models characterized candidates and implicitly anticipated the outcome, according to MIT News. For the midterms, the audit asks questions about specific House and Senate contests as well as broader voting information.

Among the questions the team wants the data to answer: whether different models give systematically different information to the same question, whether answers change over the campaign, whether a model's answer changes if it is told the user is a Democrat or a Republican, and which sources the models draw on.

Berinsky offered advice to voters. "People should be sure to remember that just because a chatbot sounds confident doesn't mean it is accurate," he said. "AI systems are designed to answer the questions that people ask them in an authoritative voice even when they are wrong, incomplete, or leaning on a weak source. These answers are a reasonable place to start, but voters shouldn't stop there."

He also said AI "can cut both ways." It can help bad actors "flood the zone" with misleading text, images and video, but recent research shows that evidence-based conversations with a chatbot can durably reduce belief in conspiracy theories. "The devil is in the details in how it is built and deployed," Berinsky said. "Independent auditing like this project can help us assess the ultimate impact of such technologies."

The dashboard presents the underlying data and some preliminary descriptive analysis, and the team plans to release the dataset to other researchers. "Our goal here is not to begin from the assumption that AI is either good or bad for elections," Podimata said.

For Massachusetts voters, official information on polling places and hours is available from the Secretary of the Commonwealth's office, not from a chatbot.

Still unknown: which chatbots show the largest differences across user identities, and whether the full analysis after the election will find patterns that favor one party.