MIT tracks how chatbots’ election answers change

Building a record rather than predicting a result

MIT researchers introduced the LLM Election Observatory on October 30 to document chatbot answers about the 2026 U.S. midterms. Adam Berinsky, Chara Podimata and Charles Stewart III lead the project.

Questions cover voting information, candidates, certification and issues. The team varies prompts, including location and political identity, then repeats the exercise every few days.

The earlier study explains the method—and its limits

The project builds on a study of the 2024 presidential election, published as a preprint in September 2025. That work tested 12 model configurations with more than 12,000 questions on a near-daily schedule from July through November 2024.

Those counts describe the earlier study, not a verified inventory of the new observatory. Its researchers queried models through programming interfaces, rather than measuring every response seen in consumer chatbot apps.

Some configurations were given search tools by the researchers. That is another reason not to treat the old results as a direct scorecard for today’s products.

The paper also cautions that its simulated identities are not representative of all the ways people describe themselves. Its fixed survey supports repeated comparisons, but cannot adapt to every unexpected answer or new event.

Presentation differences are not voter effects

Podimata describes preliminary differences in tone, cited sources and presentation in the 2026 data. The full analysis over time is not finished, and the team cannot yet establish how stable those differences are or how people react.

The earlier paper likewise explicitly leaves effects on users outside its study. An audit of generated text and an experiment measuring persuasion answer different questions.

For readers, the distinction is practical: a striking response can be worth investigating without being evidence of a broad political effect. A useful next step is research that connects documented answers to carefully measured human responses, without assuming that a change in wording means a change in belief.