MIT launches live audit of chatbot answers on U.S. midterm elections
MIT researchers have launched the LLM Election Observatory, a public dashboard that tracks in real time how popular AI chatbots answer questions about the 2026 U.S. midterm elections. The project is led by Professor Adam Berinsky, Assistant Professor Chara Podimata, and Professor Charles Stewart III, whose work spans voting and election administration, AI and misinformation, and human decision-making.
The observatory sends a fixed set of election-related prompts to different AI systems every few days. It records how answers change as the campaign develops and tests whether responses vary with a user's location, political affiliation, identity, or the way a question is phrased.
Building a record of an evolving information environment
The researchers say a central challenge is that both models and the information available to them change continually. A chatbot may provide one answer about an election at a particular point in time and a different answer weeks later, while it can be difficult to reconstruct what a voter encountered in the interim.
The dashboard is designed to create that historical record and make the collected data available to the public and other researchers. Rather than waiting for a retrospective academic analysis, users can inspect underlying responses and preliminary descriptive analyses while the election is unfolding. The team expects the dashboard to evolve alongside its data and methods.
Its questions cover election certification, voting information, prominent issues, news coverage, result predictions, and contest-specific questions about U.S. House and Senate candidates. The 2026 cycle offers a broad testing environment because state and local contests will unfold at the same time as national political narratives.
Accuracy is not the only issue
Podimata said the project is intended to assess more than whether a chatbot gives factually correct information. Models can frame the same facts differently, focus on different details, signal differing levels of uncertainty, or cite different sources. Those differences may shape how users interpret an issue even when the underlying facts are unchanged.
Early observations from the 2026 data point to variation in presentation, cited sources, and tone across the user identities supplied to the systems. The researchers have not yet completed a longitudinal analysis and do not claim to know how stable these differences will be, how they will change during the campaign, or how voters will react to them.
Private conversations, public scrutiny
Unlike discussions on major social platforms, chatbot interactions are largely invisible outside each individual conversation. Users may not know that another person asking the same question could receive a different response. The observatory seeks to make such differences visible enough to compare and scrutinize systematically.
Berinsky cautioned that an authoritative-sounding chatbot response is not necessarily accurate, complete, or based on a strong source. He described AI as capable of generating misleading material at scale, while evidence-based chatbot dialogues may also help reduce unsupported beliefs. The research does not begin from the assumption that AI is inherently beneficial or harmful to elections; it focuses on documenting what major systems actually tell people.
For organizations using chatbots in high-stakes information settings, the practical implication is to preserve outputs over time, test how prompts and user context affect answers, and require independent verification rather than treating confident language as evidence.