{"resourceId":"searchable-ai-local-issues-misinformation-midterms-gemini-6-2pct-false-20261007","versions":[{"version":"external-a66f9b4926e55c7adfb67e3e58f3c9f23f7c9fca68cfc5399e28e6008f73a95f","resource":{"id":"searchable-ai-local-issues-misinformation-midterms-gemini-6-2pct-false-20261007","title":"Searchable midterm study: 1 in 12 AI chatbot answers on local political issues contradicted official data; Gemini false in 6.2%, Google AI Mode 2.9%","organization":"Searchable Limited (AI search analytics company)","sector":"State and local government / elections and civic information","geography":"United States (Atlanta, Charlotte, Houston, Las Vegas, Portland, Maine and their states)","publishedAt":"2026-10-07 (updated)","publicationDate":"2026-10-07","eventDate":null,"sourceName":"Searchable","sourceLabel":"Independent (non-Google) commercial research report","sourceUrl":"https://www.searchable.com/data/ai-misinformation-us-midterm-elections","evidenceClass":"independent-research","outcomeClass":"cautionary","topics":["knowledge-work","governance-procurement"],"finding":"Searchable asked ChatGPT, Gemini, Google AI Mode and Perplexity 106 questions about local political issues in five U.S. metro areas and graded 7,449 answers against official government data. Overall 8.2% (610) of answers were false. Perplexity was false 22.6% of the time, Gemini 6.2%, ChatGPT 4.4% and Google AI Mode 2.9%. Gemini was most often wrong on cost of living and housing (13.0%) and schools (8.7%). Gemini read an average of 2 pages from 2 websites per answer, the fewest of the four, and drew under a third (32.2%) of its pages from government websites, the lowest share, with 15.1% from company websites. Why now: the page was updated Oct. 7, before the last google run, but the Nov. 3, 2026 midterms are three days after this edition, making its findings on Google tools newly relevant to election offices and civic-information teams.","sledRelevance":"Relevant to state and local election offices, communications teams and agencies that publish official data residents may access through Google AI products: it quantifies how often Gemini and Google AI Mode misstate local official figures and how little Gemini relies on government pages. It evaluates consumer chatbots, not Google Public Sector enterprise deployments grounded on agency data.","evidence":"Structured test: 106 questions (59 single-answer factual, 47 built from fact-checked false claims), multiple prompt forms, each asked five times in fresh sessions via anonymized logged-in accounts (not APIs); an answer counted false only when two graders agreed, with correct answers confirmed by two researchers against official sources (Census, BLS, FBI, EIA, CDC, NAEP). 43,079 pages read were classified by type. Not peer reviewed; model versions and test dates are not stated on the page.","architectureImplications":"Shows that retrieval breadth and source mix differ sharply between Gemini (few pages) and Google AI Mode; for agency deployments, grounding on authoritative datasets rather than open web retrieval is the relevant design control the study implicitly supports.","governanceImplications":"Supports agency practices such as publishing machine-readable official figures, monitoring how AI tools describe agency data, and labeling AI-assisted public answers. It does not test enterprise configurations with agency-controlled grounding.","securityPrivacyImplications":"No privacy findings; the risk is integrity of public information, including potential misinformation about local conditions near an election.","caveats":"Produced by a company that sells AI-visibility products, a commercial interest in showing chatbot errors; five metros only; model versions, dates of testing and confidence intervals not disclosed; consumer interfaces only. Published before the previous google run; included for election timing. Interpretation in role guidance is ours.","streamIds":["google"],"roles":{"sales":"Sales framing: this is cautionary third-party evidence to raise proactively with election officials and civic-information teams considering Google AI for resident questions. It supports a bounded conversation about grounding answers in official data and monitoring accuracy, since consumer Gemini misstated local figures in 6.2% of graded answers and read few government pages. It cannot support claims about the accuracy of Gemini for government services or any agency-grounded deployment, which the study did not test, nor claims that any tool is safe for unsupervised election information.","engineering":"Engineering framing: the study isolates retrieval behavior as a risk factor: Gemini averaged 2 pages from 2 websites per answer and drew 32.2% of pages from government sites. For agency resident-facing assistants, engineers should ground responses on curated official datasets, require citations, and build evaluation sets similar to Searchable's (single-answer facts plus known false claims, repeated runs, dual grading). Because model versions and test dates are undisclosed, results should be replicated on current models and the agency's own data before drawing conclusions.","delivery":"Delivery framing: an agency piloting a Google AI assistant for resident or election questions should adopt the study's method as an acceptance test: a fixed question bank drawn from its own published figures and known misinformation, asked repeatedly, graded by two reviewers. Ownership belongs to the communications or elections office with IT support. Ongoing monitoring matters because frequently updated figures, such as housing costs and school data, had the highest error rates; refresh cycles for grounding data must track publication schedules."},"roleDetails":{"sales":{"customerProblem":"Election offices and agencies need residents to receive accurate local facts, but AI chatbots, including Gemini, sometimes contradict official data.","stakeholders":["Election officials","Agency communications and public information officers","State and local CIOs","Open data program managers"],"discoveryQuestions":["Which official figures do residents most often ask about?","Is any AI assistant planned for resident or voter questions, and how will it be grounded?","How does the agency currently monitor how AI tools describe its data?","Who signs off on AI-generated public answers?"],"valueHypothesis":"A grounded, citation-required assistant with an accuracy test suite could reduce misstatements relative to consumer chatbots; this must be demonstrated on the agency's own data.","potentialEngagement":"A bounded accuracy-evaluation exercise replicating Searchable's dual-graded method on an agency's published figures before any resident-facing launch.","unsupportedClaims":["Accuracy of Gemini for government services or agency-grounded deployments","That Google AI Mode is reliable for election information","Error rates for model versions not tested","Results beyond the five metros studied"]},"engineering":{"fit":"Applies to any agency considering Gemini or Vertex AI for resident-facing information retrieval.","architectureIntegration":"Ground on authoritative agency datasets with retrieval restricted to official sources, mandatory citations, and logging of answers for review.","prerequisites":["Machine-readable official datasets with update dates","Question bank including known false claims","Two-reviewer grading process"],"constraints":["Study covers consumer interfaces only","Model versions and test dates undisclosed","Five metro areas only"],"security":"Integrity controls: source allow-lists, citation verification and answer logging to detect and correct misinformation.","proofOfValue":["Replicate the question-bank test on current Gemini and Vertex AI configurations","Measure false-answer rate on frequently updated figures","Compare grounded versus open-web retrieval error rates"]},"delivery":{"work":"Build and run a recurring accuracy test suite for any AI assistant answering resident questions about local data.","dependencies":["Curated official data","Reviewer time","Vendor configuration access"],"ownership":"Elections or communications office, with agency IT and open data teams.","skillsAdoption":"Staff need training in evaluation design and grading consistency; the study's two-grader method is a practical template.","governanceCheckpoints":["Pre-launch accuracy test","Pre-election re-test","Post-incident review of any false answer"],"acceptanceCriteria":["Proposed: false-answer rate below an agency-set threshold on the agency question bank","Proposed: 100% of answers cite an official source","Proposed: grounding data refreshed within a defined period after each official update"],"risks":["Misinformation near elections","Over-reliance on vendor accuracy claims","Stale data on frequently updated figures"]}},"retrievedAt":"2026-11-01T03:08:13Z","enrichedAt":"2026-11-01T03:13:09Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Residents relying on AI summaries, including those with limited time or language barriers, are exposed to these errors; agencies should keep authoritative plain-language sources easy to find.","procurementImplications":"Agencies can require vendors to support agency-run accuracy testing and source restrictions as contract terms.","operatingModelImplications":"Monitoring how AI tools represent agency data becomes an ongoing communications function, especially around elections.","peerReview":"unknown","sourceVerification":{"openedUrl":"https://www.searchable.com/data/ai-misinformation-us-midterm-elections","referenceExcerpt":"Perplexity gave a false answer more than 1 in 5 times (22.6%), over three times the rate of the next chatbot, Gemini (6.2%).","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}