Civic AI Audit · National Conference on Citizenship
What do AI assistants tell voters?
We ask the AI products people actually use the questions voters actually ask about the 2026 elections, every day, and keep every answer and every source they cite. Three research tracks, one growing record.
Track 1
“Candidate A or Candidate B?”
The tersest possible ask, put to every surface for every Senate, governor and battleground House race with one Democrat and one Republican, in both name orders. A judge model classifies each answer: does it pick, hedge, compare, ask what you meant, or decline?
The explorer shows what each surface does with the question, who gets picked by party, whether name order or incumbency moves the result, and every race against every surface.
Track 2
What sources do they reach for?
Every day we ask each surface to compare the candidates in a standing set of races, and record every page it cites or consults: news or campaign site, wire or advocacy group, left-leaning or right. Over time this shows what information each product is built on, and how that shifts.
Views by platform, state, race, question type, over time, and the overlap between surfaces arrive with the explorer.
Track 3
Do they get the basics right?
Our earliest experiments: simple factual questions such as who your senators are or who represents your district, graded against sourced, time-versioned ground truth by a judge model with human review.
Senate and House maps, per-surface accuracy, and the graded answers list arrive with the explorer.
The data
Everything above is built from these records. Each list can be filtered, sorted and exported.
How it works
Questions are generated from candidate rosters and asked through two channels: the official APIs, pinned to the model each product's free tier runs, and the real consumer apps driven in a browser. Judge models classify answers and sources; humans review. Detailed listings and individual answers are available to visitors who sign in. More about the instrument →