MIT researchers have launched a public dashboard tracking how AI chatbots answer questions about the 2026 US midterm elections. Early observations show that answers can differ in tone, framing and cited sources across user identities, even when the underlying facts are the same.
MIT CSAIL Watch analysis
What happened
The LLM Election Observatory, led by researchers at MIT, runs a fixed set of election questions through different chatbots every few days. The questions cover topics including election certification, candidates, predictions and news coverage. The team varies details such as a user’s location, political identity and how a question is framed, then records the answers over time.
The researchers say the public dashboard shows the collected responses and preliminary descriptive analysis, with data intended to be made available to other researchers. Their early observations suggest answers can vary in presentation, sources and tone across the identities they test. MIT News interviewed project researchers Chara Podimata and Adam Berinsky about the work.
Why it matters
A chatbot’s answer is not necessarily the same for everyone who asks the same question. Differences in emphasis, uncertainty or sources could shape what people take away, even when the underlying facts are accurate. A public, ongoing record gives researchers and voters a way to compare those answers rather than relying on one person’s chat history.
The researchers are careful about what the early results establish: they have not completed the longitudinal analysis, and do not yet know how stable the differences are or how real people respond to them. The dashboard is a way to observe the information environment, not a verdict on whether chatbots are changing votes.
Our read
This is a useful move from arguing about what chatbots might tell voters to keeping a record of what they actually say. The particularly interesting test is whether the same facts, presented with different framing or confidence, leave people with different impressions. A confident answer is not a ballot guide, as MIT political scientist Adam Berinsky reminds readers.
What to watch
- Whether differences in framing, tone and citations persist as the election approaches.
- How answers vary between chatbot models and across the tested user identities.
- When the team releases its fuller longitudinal analysis and underlying dataset.
Discussion spark: Should researchers and the public be able to compare political chatbot answers across different user profiles, or does publishing those differences risk amplifying them?
Sources and evidence
- Q&A: Hey AI, who should I vote for this year? (30 October 2026, 19:00 UTC)
Independent WittyWires tracker for public updates about MIT CSAIL. Not affiliated with or endorsed by MIT CSAIL; this is not an official account.