How accurately AI assistants answer voters

1. What a civic reading measures

Voters now ask AI assistants who is running, what they promised, when and where to vote, and who won. A civic reading puts those questions to the main assistants, in English and the languages a city's voters actually use, and checks every answer against the official record.

It answers one question: did the assistants tell voters the truth, and did they treat the candidates evenly? It does not measure which candidate AI "prefers". In an election a well-behaved assistant should not recommend anyone, so a candidate-level recommendation score would measure the wrong thing.

2. The rules

1. Score the assistants, not the candidates. The scorecard ranks ChatGPT, Gemini, Copilot and Google AI Mode on accuracy and balance.
2. Grade facts only against an official answer key: certified results, the election calendar, voter eligibility and voting locations as published by the returning officer.
3. Treat policy questions as balance checks. Both main candidates must be covered with comparable depth and accuracy. There is no official answer to grade a platform against, so policy answers are never scored as right or wrong; factual errors in them are reported.
4. Never score candidate order. Assistants list the incumbent or winner first. Order is reported, never treated as preference.
5. Ask balanced questions. Every issue question names both candidates or neither. Questions are published with the reading, and before capture for readings during a campaign.
6. Ask in the languages voters use. English plus the largest home languages in the jurisdiction, with translations checked by native speakers.
7. Respect the campaign. No new reading is published in the final 72 hours before a vote. Readings during a campaign report accuracy and balance only.
8. No political clients. The publisher and its related companies take no candidate, party, third-party campaign or lobbyist as a client in any jurisdiction a civic reading covers.

3. Readings

4. What comes next

Each future civic reading is announced on this page with its questions and capture dates before it runs. Campaign-period readings add practical voting questions (dates, eligibility, ID, where to vote) and balanced issue questions.

Civic readings follow the Trust Index methodology for capture, retention and corrections. These rules are a draft (civic rules v0.9) and will be finalised after review.