Two voters can ask the same chatbot the same question about the midterms and come away with different pictures of the race, even when every fact they are given is correct. That is the early finding from MIT's LLM Election Observatory, a public dashboard that tracks what popular AI chatbots tell people about the 2026 US elections. It matters because chatbot answers, unlike social media feeds, are private to each conversation, so nobody outside can see how they differ.
The project is run by Chara Podimata of MIT Sloan and the political scientists Adam Berinsky and Charles Stewart III. Every few days the team puts a fixed set of questions to several chatbots, covering House and Senate contests, election certification, contested issues, result predictions and news coverage, and varies the user details supplied with each question: location, party affiliation and other aspects of identity. The aim is a running record of what voters would have seen at each point in the campaign, which is otherwise impossible to reconstruct once models are updated. The team plans to release the dataset to other researchers.
In an interview with MIT News, Podimata described the first patterns in broad terms only. Across the user identities the team presents, answers differ in presentation, tone and the sources the model cites, sometimes when the underlying facts are identical. Her 2024 study of the presidential election had already found that answers drifted over the course of the campaign and shifted with the wording of the prompt, and that models differed in how they described candidates and implicitly anticipated the result.
The worry is less about outright falsehoods than about framing. Berinsky's point is that chatbots answer in a confident voice even when they are wrong or leaning on weak sources, and that a voter who wants one clear answer may not notice it was assembled from a different mix of material than a neighbour's. The same issue applies to any general assistant, including ChatGPT and the other chatbots we review, whenever users let them remember who they are. Berinsky also notes the other side: earlier research has shown that evidence-based conversations with a chatbot can durably reduce belief in conspiracy theories.
The limits are stated plainly by the researchers themselves. They have not run the full analysis, have published no effect sizes, and cannot yet say how stable the differences are or whether real people would change their views because of them. The dashboard shows preliminary descriptive data that the team expects to revise. For now this is a well-designed audit in progress, not a measured result, and the more useful output will be the dataset once the election is over.