MIT researchers introduce dashboard tracking how AI election answers vary by user

MIT researchers have introduced a public dashboard to track how AI chatbots answer questions about the 2026 U.S. midterm elections. In an October 30 MIT News interview, they described an early pattern: changing the user identity presented to a chatbot can produce differences in tone, presentation and cited sources, sometimes without changing the underlying facts.

The same questions, different user profiles

The LLM Election Observatory is led by MIT professors Adam Berinsky, Chara Podimata and Charles Stewart III. The team designed it to simulate how people across the country communicate with chatbots, rather than capture only one answer to one political question.

Every few days, the researchers put the same fixed set of election questions to different AI systems. They vary information supplied about the user, including location, political affiliation and identity. That repeated approach lets them track answers as the election approaches while comparing responses to different profiles.

The questions span election certification, divisive political issues, result predictions, news coverage and voting information. They also cover specific House and Senate contests. Podimata said the midterms offer a chance to examine many state and local races alongside national political narratives, extending her work on chatbot responses during the 2024 presidential election.

The public dashboard displays collected responses and preliminary descriptive analyses. The researchers also plan to make the dataset available for others to study. Podimata said the aim is to let people examine the information while the election unfolds, rather than wait months for an academic paper.

Differences in framing, emphasis, uncertainty, and cited sources can matter enormously, even in the absence of outright misinformation.

Chara Podimata, speaking to MIT News

Correct facts can still produce different impressions

Podimata’s concern goes beyond whether an answer contains a false claim. A chatbot might emphasize one part of an issue, express more doubt about another, or choose different sources. She warned that two users could leave with different perceptions of the same facts without realizing their conversations had diverged.

The early observations are not a completed assessment of that risk. The team has not finished its longitudinal analysis—tracking patterns over time—and cannot yet say whether the differences persist, how they shift during the campaign, or how real people interpret and react to them.

The audit is designed to investigate several distinct questions: whether models give systematically different information, whether their answers change during the campaign, whether a user’s self-description changes a response, and which sources the systems draw on. Those are research questions, not findings that the dashboard has already settled.

Preserving answers that otherwise disappear

Another purpose is historical. Podimata said model updates, changing political information and differences in how questions are asked make this information environment difficult to reconstruct. If a system answers differently several weeks later, researchers may struggle to recover what voters would have seen at the earlier moment.

She contrasted chatbot conversations with social media, where researchers could at least partly observe information circulating across groups and accounts. Individual chatbot exchanges are largely hidden from outsiders. Collecting and sharing responses is intended to make those differences visible enough for public scrutiny and further research.

For voters, Berinsky offered a narrower, immediate caution: a confident answer is not necessarily an accurate one. Chatbots can sound authoritative while being wrong, incomplete or dependent on weak sources. He described their answers as a reasonable starting point for election research, but urged voters not to stop there.