Millions of voters now ask AI chatbots about politics. We are tracking what they hear back.
Caucus AI is building a public record of what AI tools tell Americans about candidates and elections, so we can spot patterns, catch mistakes, and understand how AI is shaping political information.
Ask ChatGPT "Who's running for Congress in my district?" Then ask Gemini and Grok. Are the answers the same? Are any of them right?
Every week, we put the same questions to the major AI chatbots and summaries and show their answers side by side, along with the sources they cite.
Why does this matter?
AI is replacing search. Millions of Americans now ask Generative AI chatbots the same questions they used to type into Google or ask a friend. 45% of Americans report having heavy exposure to AI-generated search summaries and 74% of AI users reported using AI tools for search and internet browsing. Reliance on AI looks like it will be widespread among future voters as well: 64% of US teenagers use chatbots, with 28% reporting daily use.
AI is becoming a source for news and political information. While relatively few Americans currently get their news from AI chatbots, AI-generated search summaries and searches for political factual questions are on the rise. In the case of Google's AI overviews, the AI response is often the single source of information, as users presented with AI search summaries are significantly less likely to click on primary source links. Users may end up relying on single, authoritative answers that are often overly certain, lacking nuance, or simply incorrect. Trust in the accuracy of these tools is low, reflecting awareness of some of these issues.
How politics is different
AI companies prioritize neutrality. Most AI companies, including Google, Anthropic, and OpenAI, currently take significant pains to make their flagship models "politically even-handed". This often results in bland, almost painfully cautious responses that are in stark contrast to model responses on other topics, like restaurant recommendations.
But neutrality is difficult in practice. AI systems are trained on huge amounts of human-generated text that bake in existing societal biases and points of view, and can exhibit subtle bias or be easily prompted to give biased answers. Post-training adjustments, like instructions to be neutral or instructions to correct perceived bias, can result in unexpected behavior. At its most extreme, many of us witnessed the transformation of Grok into "Mecha-Hitler", the result of xAI's finetuning efforts to make Grok slightly more conservative.
The Trump Administration is also working to define neutrality in AI, raising questions about whether government pressure will compromise efforts to achieve unbiasedness.
When AI chatbots try to persuade, they're effective. Research has shown that AI chatbots can be hugely persuasive when they are instructed to be, with effects significantly larger than traditional persuasion tactics, like television ads. If flagship AI models were to become partisan, they could have a large electoral effect.
AI companies care about political outcomes. AI companies have a significant stake in outcomes of American elections and are spending accordingly. AI policy, regulation, and government contracts are all issues that will only become more salient in American politics in coming months and years. There is little transparency on how these companies create and fine-tune their models and few guardrails to ensure these models cannot be used to further companies' political goals.
Who we are
Meg Schwenzfeier is a data scientist in Democratic politics, most recently Chief Analytics Officer on the Harris-Walz campaign and before that head of Analytics at the DSCC and of data science on the Biden-Harris 2020 campaign.
Jack Welty is an engineer in Democratic politics, most recently Deputy Chief Analytics Officer on the Harris-Walz campaign and before that at the DSCC, the White House, and the Biden-Harris 2020 campaign.
Methodology
This tool asks multiple AI models the exact same questions about political candidates and races. We store every response, extract the sources each model cites, and compare and contrast answers across model providers. With each query, we allow the model to use web search tools, to better mimic how AI chatbots and AI-enhanced search behave for real users. This creates a database that makes it possible to:
- See what AI models tell users in response to questions about candidates and districts
- Understand what sources models cite for political information
- Spot factual errors, out-of-date information, and hallucinations
- Compare how various models answer the same question
- Track how model responses change from week to week
FAQ
Q: Why don't you have all candidates for all races?
A: We cover every 2026 Senate, Governor, and U.S. House race. Please flag for us if we're missing anyone you want to see!
Q: Will the model responses be updated?
A: We refresh the model responses every week.
Q: How did you pick which models to query?
A: Our goal is to simulate the chat experience of an average voter as closely as possible, so we query the default models behind the most commonly-used chat apps: ChatGPT, Gemini, Grok, and Claude, along with Google AI Overviews. When a provider changes its default, we do our best to follow it.
Q: What should I do if a model provides a factually inaccurate response?
A: As part of our research, we are interested in collecting a dataset of inaccurate LLM responses. As you browse the site, please send any factual inaccuracies you spot to [email protected]. We plan to put out additional research on patterns we find in these inaccuracies, and are seeking out partners to test mitigation strategies.
Contact
Interested in learning more or collaborating with us on this research? We'd love to hear from you! Reach out at [email protected].