{"resourceId":"honeyb-ai-midterms-tracker-gemini-voter-answers-accuracy-20261029","versions":[{"version":"external-3d5fb768bae8f8f9ef304bd3db861b4ed760dda9a44b7cccf02fcecc0fd35541","resource":{"id":"honeyb-ai-midterms-tracker-gemini-voter-answers-accuracy-20261029","title":"Honeyb AI Midterms Tracker (updated Oct. 29): Gemini 2.5 Flash answers to state voter-rule questions matched official rules 94% of the time, with 14 wrong, 2 outdated and 7 no-answer results","organization":"Honeyb (commercial AI-search analytics firm)","sector":"State government (election administration and voter information)","geography":"United States (50 states and DC)","publishedAt":"Live tracker; measuring since 2026-09-14, page updated 2026-10-29","publicationDate":"2026-10-29","eventDate":"2026-11-03","sourceName":"Honeyb","sourceLabel":"Independent commercial tracker (non-academic; methodology published on page)","sourceUrl":"https://www.honeyb.ai/live-data/us-midterms-2026","evidenceClass":"independent-research","outcomeClass":"mixed","topics":["knowledge-work","governance-procurement","accessibility-workforce"],"finding":"The tracker asks ChatGPT and Gemini (gemini-2.5-flash, web search on) voter questions on registration, ID, mail ballots, early voting, polling places, same-day registration and Senate/governor candidates, and checks answers against fact tables from state election offices cross-checked with NCSL and Ballotpedia. As of Oct. 29, Gemini scored 94% correct with 14 wrong, 2 outdated and 7 no-answer results, versus ChatGPT 89%. Gemini cited official election sites in 36% of its 3,559 citations; its procedural answers were 96% correct when citing official sources and 94% when not. Senate-race questions were hardest across engines (84%).","sledRelevance":"Five days before the Nov. 3 election, state and county election officials can see how often Gemini relays their rules accurately and where it errs (e.g., naming candidates not on the ballot). This informs voter-education messaging, official-source publishing and whether to rely on consumer AI as a channel.","evidence":"Daily automated measurement in 14 battleground states and alternate-Sunday checks of all states; a fixed model extracts claims and code compares them with a fact table; reasons published per verdict. Page header cites 1,283 answers checked while the summary reports 92% of 1,106 checked answers across engines; the discrepancy is unexplained.","architectureImplications":"AI assistants with web search pull from official sites, Wikipedia and news; election offices that publish clear, structured, current rules pages improve the chance of correct AI answers.","governanceImplications":"Election offices may want to monitor AI-assistant answers about their jurisdiction and issue corrections or guidance directing voters to official sources.","securityPrivacyImplications":"Wrong or outdated answers create misinformation risk around deadlines and eligibility; the tracker does not test adversarial manipulation.","caveats":"Commercial, non-peer-reviewed tracker run by a firm that sells AI search visibility services; consumer Gemini, not a government deployment; English-only from a U.S. location; live page whose figures change daily; count discrepancy between header and summary.","streamIds":["google"],"roles":{"sales":"Frame this as voter-information quality, not a Gemini sale. State and county election offices care whether AI assistants repeat their rules correctly; this tracker shows Gemini is mostly right but still names candidates not on the ballot and sometimes gives no answer. A bounded offer would help an office structure its official rules content and monitor AI answers. Do not imply that a Google product deployment fixes consumer-assistant errors.","engineering":"The technical lever is the official source. Gemini's procedural answers were correct 96% of the time when they cited official sites, yet only 36% of its citations were official election sites. Engineers can help election offices publish clear, current, machine-readable rules pages and test what assistants return for their jurisdiction using a method similar to the tracker's claim extraction against a fact table.","delivery":"Package as a short pre-election readiness task: build a jurisdiction fact table, run a repeatable set of voter questions against major assistants, log discrepancies and push corrections through official channels and voter-education messaging. Assign a communications owner and repeat checks as deadlines pass. Results here are a commercial tracker's, so validate locally before acting. Prioritize the hardest categories, such as Senate-race candidate questions."},"roleDetails":{"sales":{"customerProblem":"Election offices cannot see whether AI assistants give voters accurate rules for their jurisdiction.","stakeholders":["Secretary of State elections division","County election directors","Public information officers"],"discoveryQuestions":["Does the office monitor what AI assistants say about its rules?","Are official rules pages current and structured?","How would the office respond to a widely repeated AI error?"],"valueHypothesis":"Monitoring and better official content could reduce voter confusion from AI answers; impact is not measured.","potentialEngagement":"Pre-election AI-answer audit and official-content improvement sprint.","unsupportedClaims":["Gemini accuracy for any specific state at a given time","Effect of content changes on AI answers","Accuracy in languages other than English"]},"engineering":{"fit":"Relevant to state and county election offices with public rules pages.","architectureIntegration":"No integration with Google systems; relies on public web content consumed by assistants with search.","prerequisites":["Authoritative jurisdiction fact table","Question set mirroring voter queries"],"constraints":["Assistant outputs change daily","No control over third-party models"],"security":"Watch for data voids that adversaries could exploit; monitor for manipulated content.","proofOfValue":["Repeatable question run with documented discrepancies","Track change in accuracy after content updates"]},"delivery":{"work":"Run AI-answer checks against a jurisdiction fact table and remediate official content.","dependencies":["Elections staff to confirm rules","Communications team"],"ownership":"Election office public information officer.","skillsAdoption":"Train staff on running and interpreting checks.","governanceCheckpoints":["Fact table sign-off by elections director","Review before public corrections"],"acceptanceCriteria":["Proposed: discrepancy log maintained through Election Day","Proposed: official pages updated within set time of rule changes"],"risks":["Over-reliance on a commercial tracker","Rapidly changing assistant behavior"]}},"retrievedAt":"2026-10-30T03:11:12Z","enrichedAt":"2026-10-30T03:18:56Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"English-only testing leaves language-access accuracy unmeasured; election staff may need to address AI-sourced confusion at help lines.","procurementImplications":"Limited direct procurement relevance; informs decisions on whether to invest in structured official content and monitoring.","operatingModelImplications":"Election communications teams may add AI-answer monitoring to pre-election readiness checklists.","sourceVerification":{"openedUrl":"https://www.honeyb.ai/live-data/us-midterms-2026","referenceExcerpt":"Each answer is checked against the state's official rules.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}