MIT researchers are building a public record of how AI chatbots discuss the 2026 U.S. midterm elections. In an October 30 interview, the team described the LLM Election Observatory, which repeatedly tests a fixed set of questions rather than relying on isolated screenshots.
The project examines responses about contests, election administration and political issues. Questions are rerun every few days with variations in user descriptions, including location and party affiliation. Adam Berinsky, Chara Podimata and Charles Stewart III lead the work.
Podimata reported preliminary differences in tone, framing and cited sources across user identities, sometimes alongside unchanged underlying facts. She emphasized that the full longitudinal analysis was still unfinished. The team could not yet determine how stable these patterns were or how actual people would interpret and react to them.
The dashboard provides underlying responses and descriptive analysis while the election unfolds. It is an observational audit of system outputs, not evidence that a particular chatbot has changed votes.
NAOAT analysis
The project exposes a measurement problem that also matters outside politics. A model can preserve the factual core of an answer while changing the emphasis enough to alter the impression a reader receives. A simple correct-or-incorrect score would miss that difference.
Repeating the same questions creates a reference against which changes become visible. Without a dated response record, it is difficult to distinguish a persistent pattern from a temporary output, a changed model or a different prompt. The observatory’s value lies partly in retaining that history for later scrutiny.
For the technology sector, the distinction between output measurement and human impact is especially consequential. The former can identify behavior worth investigating; the latter requires evidence about people. Keeping those claims separate makes an audit more credible and prevents preliminary observations from becoming unsupported conclusions about persuasion.