Every few days, MIT researchers put the same election questions to popular AI chatbots and record what comes back. Their LLM Election Observatory is designed to show how answers about the 2026 U.S. midterm elections change over time—and how they may differ depending on who is asking.
More people are turning to chatbots to find and make sense of political information. But an answer can shift as an AI system is updated, new information becomes available, or a user phrases a question differently. Without a record, it can be hard to know what voters might have been told at a particular point in an election campaign. The MIT dashboard aims to build that record as the midterms approach and make its data available to researchers and the public.
The project is led by Adam Berinsky, a professor of political science; Chara Podimata, an assistant professor of operations research and statistics; and Charles Stewart III, a professor. Their team asks different AI systems a fixed set of questions on topics such as election certification, contentious issues, predictions about results and news coverage. It repeats the questions every few days, testing how answers vary with factors such as a user's location, political party affiliation and identity.
The observatory is meant to track more than outright false claims. Podimata says researchers also need to examine how chatbots frame issues, what details they emphasize, how much uncertainty they express and which sources they cite. Those choices can influence how people understand political questions even when an answer does not contain a clear factual error.
The work builds on Podimata's research during the 2024 presidential election. That work found meaningful differences in AI responses over time and depending on how prompts were framed. It also found differences in how models described candidates and implicitly anticipated the election outcome. The 2026 midterms offer a chance to extend that research, while the observatory's repeated checks can help show whether answers are changing as the election gets closer.
The dashboard does not claim that a chatbot will persuade a voter to change parties. Instead, it gives researchers a way to examine what political information these systems provide and how that information may shape public discussion. By preserving snapshots of chatbot answers, the team hopes to make changes visible rather than leave them difficult to reconstruct after the fact.
As campaigns unfold, the observatory will offer a continuing record of how AI systems respond to the same questions under different conditions. That record could help voters, researchers and the public better understand a growing part of the political information landscape—and ask not only whether a chatbot's answer is accurate, but also how it guides attention and understanding.