As voters turn to chatbots for political information, a new MIT project is tracking what those systems tell them about the 2026 U.S. midterm elections. The LLM Election Observatory collects responses to the same questions over time and under different user circumstances, making changes in this personalized information environment easier to study.
A changing record of election answers
AI systems and the political information available to them can change during an election cycle. A chatbot’s answer may also depend on how someone phrases a question or what they say about themselves. That makes a single snapshot an incomplete record of what voters encounter.
Assistant Professor Chara Podimata says the project aims to document chatbot responses systematically and make the resulting data available to researchers and the public. Without such a record, it can be difficult to reconstruct what a system told users at an earlier point if its answer changes weeks later.
The observatory runs a fixed set of questions through different AI systems every few days. The questions cover topics such as election certification, contentious issues, predictions about results, and news coverage. The team also varies factors including location, political party affiliation, and identity to examine how responses differ.
Accuracy is only part of the picture
Checking whether a chatbot gets facts right is essential, but the researchers say it does not capture the full influence of an answer. Two responses can be factually accurate and still leave readers with different impressions because they frame the issue differently, highlight different details, express different levels of uncertainty, or cite different sources.
Those choices can affect how people understand political questions without amounting to outright misinformation. The observatory is therefore designed to help researchers examine not only what information chatbots provide, but also how they present it and how that presentation changes.
Podimata’s earlier work during the 2024 presidential election found meaningful variation in AI models’ responses over time and across prompt framings. It also found differences in how models described candidates and implicitly anticipated the election outcome. The 2026 project extends that line of study to the midterms.
One election, many different information environments
The midterms involve numerous state and local contests as well as broader national political narratives. The team’s audit survey includes questions specific to U.S. House and Senate candidates, alongside broader questions about voting information and the election information environment.
Individual chatbot conversations can create distinct “microcosms” of political information. A person may not know that a friend asking about the same issue receives a different answer from a preferred chatbot. Even when both answers are accurate, one may stress different aspects of the issue or rely on different sources, leading each person to form a different view of the underlying facts.
This visibility problem has echoes of concerns about social media. Researchers could observe at least some information circulating through online groups and communities. Chatbot exchanges, by contrast, are largely private to each conversation, making it harder to compare what different people see or follow changes over time.
Making chatbot responses open to scrutiny
By collecting and sharing responses, the LLM Election Observatory aims to make those otherwise hidden differences more observable. Its ongoing comparisons can help researchers ask whether models consistently give different information, whether answers shift during the campaign, and whether a response changes with the wording of a prompt or details about the user.
For voters, the project’s central implication is that a chatbot answer is one account of an election issue, shaped by the system and the interaction. A confident or well-framed response does not, by itself, reveal whether another user was shown different emphasis, uncertainty, or sources. Tracking these patterns gives the public and researchers a way to examine how AI systems contribute to political discourse as the election approaches.