{"resourceId":"wilhelmshaven-city-chatbot-nwz-test-wrong-answers-20261104","versions":[{"version":"external-040071347ef2648830802d66ed2414ddcfa62c07b401366d811d906d465a55d0","resource":{"id":"wilhelmshaven-city-chatbot-nwz-test-wrong-answers-20261104","title":"Nordwest-Zeitung tests Wilhelmshaven's new city chatbot and finds confident wrong answers on election results, officials and an unresolved unexploded-ordnance site","organization":"City of Wilhelmshaven, Germany (tested by Nordwest-Zeitung)","sector":"Municipal AI website assistant","geography":"Wilhelmshaven, Germany","publishedAt":"2026-11-04","publicationDate":"2026-11-04","eventDate":"2026-10-31","sourceName":"Nordwest-Zeitung (NWZ)","sourceLabel":"NWZ: Chatbot der Stadt Wilhelmshaven im Test: KI-Tool gibt falsche Antworten","sourceUrl":"https://www.nwzonline.de/wilhelmshaven/chatbot-der-stadt-wilhelmshaven-im-test-ki-tool-gibt-falsche-antworten_a_4,3,2259962026.html","evidenceClass":"independent-reporting","outcomeClass":"cautionary","topics":["developers-agents","governance-procurement","operating-model"],"finding":"A Nordwest-Zeitung reporter tested Wilhelmshaven's new AI chatbot, launched in October with a claim of over 100 languages. Asked on Oct. 31, it did not know about a public inauguration ceremony the city had announced on its own site on Oct. 30. It treated the Sept. 13 local election as a future event, then gave a wrong 'official final result' order, and in another session said results were unpublished. It misassigned council group spokespeople, varying across sessions, and said investigations of three suspected unexploded-ordnance sites were closed when one remains open. The reporter notes it also gave several helpful answers and calls the test a sample, not a full evaluation.","sledRelevance":"Independent follow-up on a launch the archive covered (local-government, Oct. 19): a concrete test showing stale content and confident errors, including a safety-related one. Useful evidence for acceptance testing of U.S. city chatbots.","evidence":"Journalist's hands-on test with specific questions and checked answers; small, non-systematic sample.","architectureImplications":"Errors suggest the bot does not reliably index the city's newest pages and answers inconsistently across sessions.","governanceImplications":"Safety-related answers (unexploded ordnance) should be restricted or escalated; the bot presented errors with authoritative wording.","securityPrivacyImplications":"None material in the report.","caveats":"Single reporter, handful of questions, no systematic scoring; non-US context. City response not reported.","streamIds":["citizen-engagement"],"roles":{"sales":"Independent tests of city chatbots are now appearing in local news, and they hit credibility hard. This NWZ test is a small sample but documents specific errors on elections and public safety. Discovery should ask how a city would know its bot is wrong, how fast new pages are indexed, and which topics should never be answered by AI. A bounded engagement is a pre-launch and quarterly answer audit. The source does not show the bot fails overall.","engineering":"Engineering fit centers on retrieval freshness and answer consistency. The errors point to delayed indexing of new pages and non-deterministic answers. Teams should confirm crawl frequency, ranking of recent content, and topic blocklists for safety and elections. Proof of value should replay a fixed question set across sessions, test answers on pages published within 24 hours, and check escalation for safety questions.","delivery":"Delivery work is post-launch monitoring. A content owner must run regular answer audits, route sensitive topics to staff, and make the bot visible and labeled as AI. The acceptance criteria below are proposals, not observed results. Cities should publish how residents can report wrong answers and how quickly corrections are made."},"roleDetails":{"sales":{"customerProblem":"City chatbots give confident wrong answers that residents cannot detect.","stakeholders":["City web and communications team","Elections office","Public safety office","Vendor"],"discoveryQuestions":["How quickly are new city pages indexed?","Which topics are blocked or escalated?","How are wrong answers reported and fixed?"],"valueHypothesis":"Regular answer audits can catch stale or wrong answers before they spread; the test shows the risk, not a fix.","potentialEngagement":"A pre-launch and quarterly chatbot answer audit on priority topics.","unsupportedClaims":["Overall failure rate","Vendor-wide quality","Resident harm"]},"engineering":{"fit":"Applies to any city website assistant built on site content.","architectureIntegration":"Website retrieval assistant behind a 'Moin' button; freshness and consistency are the weak points.","prerequisites":["Fixed test question set","Crawl schedule","Topic blocklist"],"constraints":["Small journalist sample","Non-US context","Vendor not named"],"security":"Restrict election and safety answers to verified pages or staff escalation.","proofOfValue":["Cross-session consistency replay","Fresh-page answer test","Safety-question escalation check"]},"delivery":{"work":"Monitor and correct chatbot answers after launch.","dependencies":["Content owners by department","Vendor indexing","Feedback channel"],"ownership":"City web and communications team.","skillsAdoption":"Staff need a routine for reviewing logs and correcting content.","governanceCheckpoints":["Pre-launch audit","Election-period review","Quarterly answer audit"],"acceptanceCriteria":["Proposed: new pages answerable within 24 hours","Proposed: zero unescalated answers on safety topics in audits"],"risks":["Confident errors on safety","Loss of public trust","Inconsistent answers across sessions"]}},"retrievedAt":"2026-11-05T04:02:22Z","enrichedAt":"2026-11-05T04:03:55Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Bot is hidden behind a small 'Moin' button and not findable via site search, limiting discoverability.","procurementImplications":"Supports including content-freshness and consistency tests in acceptance criteria; vendor not named in the article.","operatingModelImplications":"Shows the need for an owner who monitors answers after launch, especially on elections and safety topics.","sourceVerification":{"openedUrl":"https://www.nwzonline.de/wilhelmshaven/chatbot-der-stadt-wilhelmshaven-im-test-ki-tool-gibt-falsche-antworten_a_4,3,2259962026.html","referenceExcerpt":"trägt dabei absoluten Unsinn mit derselben Höflichkeit und Sicherheit vor wie Zutreffendes","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}