What are AI chatbots telling voters? MIT has logged 3.75 million answers—but not voter effects
A live MIT audit repeatedly asks election questions across model, search and identity settings. The scale is unusual, but the team has not yet completed its longitudinal analysis or tested how real voters respond.
The short version
- What happened
- MIT's live release contains 3,748,432 responses gathered on 16 dates from 29 August to 28 October across 564 election questions, 13 model/search configurations and 35 prompt variants.
- Who it affects
- Voters may receive different framing, tone and sources for the same political question without knowing that another model, prompt or profile would produce a different answer. The dashboard can make that variation inspectable, but people still need official election information and the original sources behind consequential claims.
- What remains unproven
- Ongoing institute research dashboard
Keep in mind
Ongoing institute research dashboard
Not peer reviewed
3,748,432 generated responses, not people
16 irregular collection dates
API access may differ from consumer interfaces
Related search-on and search-off configurations counted separately
Preliminary analysis
No complete accuracy denominator
No voter-behaviour outcome
Funding and conflicts not disclosed on the reviewed pages
Source analysis · 3 linked sources. Primary analysed source date: 30 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
The dataset is large, but its purpose is observation rather than persuasion testing
MIT researchers have released a live record of what major AI systems say when asked about the 2026 United States midterm elections. The 30 October data release contains 3,748,432 stored responses from 16 collection dates between 29 August and 28 October. It covers 564 questions derived from 62 templates, 13 model-and-search configurations and 35 prompt variants. Topics include candidates, voting processes, election integrity, policy issues, predictions, news and possible consequences of the result.
That scale makes the project useful for tracking an information environment that is otherwise difficult to reconstruct after models change. It does not make the responses representative of voters or chatbot use. The unit is a generated answer, not a person. Repeated outputs from related configurations contribute separately, and the dashboard describes some model labels as search-on and search-off versions of the same underlying product.[1][2][3]
The audit changes one factor at a time—and still differs from ordinary chatbot use
The team sends a fixed bank of questions to the default free-plan models it selected, using default temperature through provider APIs. It collects answers with and without web search, while Google AI Overview is treated as a separate search product. Prompt variants tell the system that the questioner has a particular political affiliation, race or ethnicity, gender identity or location. Repeating the runs every few days lets researchers compare model, prompt, search and time effects against a common question.
This design supports systematic comparison, but it is not the same as observing millions of natural conversations. API behaviour can differ from consumer interfaces, the chosen prompts cannot capture every way people ask political questions, and model versions can change. Collection cadence also varies with the project's budget. The 16 dates are irregular, with 217,140 to 256,620 responses stored per date, and two configurations appear on only seven of the 16 days.[1][2]
The distinction that matters
What the 3.75 million-response release actually contains
Four denominators describe the audit; none is a count of voters or unique chatbot users.
Questions
564 questions from 62 templates
Candidates, processes, issues, predictions, news and integrity
Configurations
13 model/search settings
Includes search-on and search-off versions plus Google AI Overview
Prompt variants
35 collected identity or instruction settings
Tests party, race and ethnicity, gender, location and confidence wording
Collection
16 dates; 3,748,432 responses
A repeated output archive—not 3.75 million people
Early differences are visible; voter consequences remain unmeasured
The researchers say preliminary 2026 data show differences in how answers are presented, which sources are cited and the tone used for different attributed identities, sometimes when the underlying facts are similar. They have not yet reported a complete longitudinal estimate of how stable those patterns are. The public release also does not supply one audited accuracy rate across all 3.75 million answers, and its similarity pipeline embeds 3,661,436 responses while separately recording blank and absent outputs.
For voters, journalists and election officials, the practical value is an inspectable record: a confident answer can be compared with other models, dates and prompt settings rather than treated as neutral truth. The missing step is human impact. Nothing in the release shows that a chatbot changed a vote, corrected misinformation, deepened polarisation or improved understanding. The reviewed pages also do not disclose project funding or conflicts, and the work has not been peer reviewed. Confidence would rise with prespecified longitudinal analyses, transparent accuracy coding, independent replication and experiments measuring what people believe or do after seeing different answers.[1][3]