View races

Frequently asked questions

How it works

  1. 1Ask

    We ask a question like 'Who should I vote for?' with an insistent prompt asking the model to provide a single top choice.

  2. 2Vary the persona

    For races with candidates from multiple parties, we repeat the question under four voter identities: no persona, liberal, conservative, and independent.

  3. 3Measure

    We run each prompt many times, classify which candidate (if any) the model recommends, record search results used in each answer, and tally the results into the percentages you see on each race page.

What is this project?

This is a study of political influence in mainstream AI assistants. We ask leading AI models the same kind of question a real, undecided voter might ask. Then we measure which candidates get recommended, and which websites are influencing these answers.

What exactly are you measuring?

For each race we measure the share of answers that recommend each candidate, broken down by AI model and by voter persona. The headline number for a candidate is the percent of a model's answers that recommended them in that slice. We also track how often models refuse to answer, stay vague, or recommend someone outside the field.


We also keep track of all the sources that each model sees when it searches the web. These are categorized by their political leaning and by what 'kind' of source they are (News, Social Media, Government, etc.)

What are the four voter personas?

We ask each question four different ways. The only thing that changes between them is the self-identified voter identity attached to the prompt.

  • No persona — No stated identity is added. This is the neutral baseline.
  • Liberal — We add a section self-identifying as a liberal voter.
  • Conservative — We add a section self-identifying as a conservative voter.
  • Independent — We add a section self-identifying as an independent voter.

For races that include candidates from multiple parties, comparing the personas shows whether a model gives everyone the same answer or tailors its recommendation to the politics it infers from the asker.

Which AI models do you test?

With an insistent prompt, we are able to consistently get candidate recommendations from ChatGPT and Grok.

On every race page you can click on ChatGPT or Grok to see how their recommendations differ, and which sources they see when they search the web.

A note on refusals: Claude, Google AI Overview, Gemini, and Google AI Mode declined to answer every voting and election question we asked, so they produced no recommendations to measure and aren’t included in the results.

How do you turn a chatbot's answer into a percentage?

We run each prompt many times, then record its candidate choice (or a refusal to answer). Each answer can be recorded for only one candidate, so the percentages you see for a single model in a single race will add up to 100%. A candidate's percentage is simply the count of answers that recommended them divided by the total answers in that slice.

What are the websites listed as 'sources used' by each model?

These AI models search the web. We collect the links they use as sources and show the domains and types of pages that appear most often for each slice. This hints at where a model's recommendation is coming from, and how that source mix shifts across personas and models.

What is the citation rate next to each source?

Wherever we list sources, the percentage next to a source is its citation rate. This measures how often that source was retrieved when the model was writing its answer.

For example, if a model answered a race’s question 100 times for liberal voters and 14 of those answers cited calmatters.org, that source’s citation rate in the slice is 14%.

A single answer usually cites several sources, so citation rates don’t add up to 100%. On a candidate’s drill-down view, the rate is measured against just the answers that recommended that candidate.

Why does this matter?

AI assistants are quickly becoming the default way people research candidates. If a model has inherent preferences for certain types of candidates, or if it sees certain types of sources more often when searching the web, that can shape political opinions at scale. This project makes those patterns measurable and comparable.

What this is not

This is not an endorsement of any candidate, and it's not a prediction of who will win. It's a measurement of what AI models say, and the sources that they read. Model behavior also changes over time, so results reflect the period in which the prompts were run.

Ready to see the results? Explore the featured races and the map.

View the races