MIT researchers launch dashboard to track AI chatbot responses to US election questions

Researchers at the Massachusetts Institute of Technology (MIT) have developed a public dashboard to monitor how artificial intelligence chatbots respond to questions about the 2026 US midterm elections.

Called the LLM Election Observatory, the project tracks responses from popular AI systems to a fixed set of election-related questions. The researchers are examining whether answers differ across chatbots and how they change depending on a user's location, political affiliation, identity and the way a question is framed.

The project is led by MIT professors Adam Berinsky and Charles Stewart III, along with assistant professor Chara Podimata. It aims to document the political information people receive through AI systems and examine how these responses could shape public understanding of election issues.

The team runs the same questions through different chatbots every few days, building a record of their responses as the election approaches. Topics include election certification, political issues, election-related news coverage and predictions about results.

AI answers may differ even when questions are identical

The researchers are examining whether chatbots provide different information to users asking the same question, including when users identify themselves as Democrats or Republicans.

According to Podimata, responses can differ in how they frame facts, which issues they emphasise, the level of uncertainty they express and the sources they cite. Such differences may influence how users understand an issue even when the underlying information is factually accurate.

The project builds on Podimata's research during the 2024 US presidential election, which found variations in AI responses over time and across different prompt framings. The new observatory expands that work to include questions about individual House and Senate contests, as well as broader aspects of the election information environment.

The public dashboard displays collected responses and preliminary descriptive analysis. The researchers also intend to make the dataset available to other researchers for further study.

Researchers caution against relying on confident AI answers

The team has identified some early differences in the tone, presentation and sources used by chatbots across different user identities. However, the researchers said a full longitudinal analysis has not yet been completed. They have not established how stable these differences are over time or how people interpret and respond to them.

Berinsky cautioned voters against treating a chatbot's confident tone as proof of accuracy. He said AI-generated answers can be wrong, incomplete or based on weak sources, and should be treated as a starting point rather than the final word when looking for election information.

The researchers also noted that AI can be used both to generate misleading content and to help challenge misinformation through evidence-based conversations. The observatory is intended to provide an independent record that can help assess how these systems are being used in the election information environment.

The team said its objective is not to assume that AI is inherently beneficial or harmful to elections, but to examine what information major AI systems provide as an election unfolds.