The Pulse
MIT Builds Dashboard to Track AI Answers on the 2026 Elections
MIT researchers launched the LLM Election Observatory to record how popular AI chatbots answer questions about the 2026 U.S. midterm elections. Early observations show differences in tone, framing and cited sources across user profiles, but
AI.info Team · Written with AI · fact-checked ·
Every few days, MIT researchers run the same election questions through popular AI chatbots and record what comes back. Their new public dashboard, the LLM Election Observatory, tracks how those responses vary across models, user profiles and time as the 2026 U.S. midterm elections approach.
One set of questions, repeated over time
The project is led by MIT professors Adam Berinsky and Charles Stewart III and Assistant Professor Chara Podimata. Its researchers ask about election certification, candidates, predictions, news coverage and other topics, then compare chatbot replies under different descriptions of a user’s location, political affiliation and identity. The repeated runs are intended to capture changes that a single snapshot would miss.
MIT says the public dashboard presents the collected responses alongside preliminary descriptive analyses. The team also plans to make the dataset available to other researchers. The LLM Election Observatory is designed to let people outside academia inspect the material while the campaign is under way, rather than waiting for a later research paper.
That record matters because chatbot answers can change as models are updated and political information shifts. A response seen today may be difficult to reconstruct weeks later, MIT researchers say. Their project aims to preserve a record of what systems provided at different points in the election cycle.
Differences may go beyond factual accuracy
Early observations show variation in how answers are presented, which sources models cite and the tone they use across the user identities tested. Podimata said those differences sometimes appear even when the underlying facts are the same. The researchers are also comparing whether answers change with question wording, user descriptions and the chatbot itself.
Those distinctions could shape what users take away from an answer without the system stating a plainly false claim. A chatbot might emphasize one aspect of an issue, express more or less uncertainty, or rely on different sources. The dashboard makes those differences available for comparison, but its early patterns do not establish how people interpret them.
“People should be sure to remember that just because a chatbot sounds confident doesn’t mean it is accurate,” said Adam Berinsky, Mitsui Professor and professor in MIT’s Department of Political Science.
Researchers have not measured voter reactions
MIT’s researchers have not completed the full longitudinal analysis. Podimata said they cannot yet determine how stable the observed differences are, how they shift over the course of the election, or how a real person would respond to them. The dashboard is therefore a record of chatbot output, not evidence that the systems have changed anyone’s vote.
The work builds on Podimata’s research into chatbot responses during the 2024 presidential election, which found variation over time and across prompt framings. For 2026, the team has expanded its questions to cover simultaneous House and Senate contests, as well as broader election information. The researchers say their aim is to document what major AI systems tell people, not to assume in advance that chatbots help or harm elections.
The next test is whether patterns persist
The project leaves two practical questions open: whether differences in tone, framing and sources hold as the campaign develops, and whether those differences affect people who encounter them. MIT’s team says future analyses will address those issues. For now, its dashboard offers a continuing record of the answers researchers have collected, while the effect on voters remains unmeasured.