{"resourceId":"korea-gov-ai-chatbots-conflicting-wrong-answers-20261026","versions":[{"version":"external-33cbf2bf2d006145c848e723dbd9db11dee74f0eb222a757e4d4419d4b464a51","resource":{"id":"korea-gov-ai-chatbots-conflicting-wrong-answers-20261026","title":"Seoul Economic Daily test finds Korea's AI Government24 chatbot gave outdated tax rules marked \"high confidence\" and conflicted with other ministry bots","organization":"Seoul Economic Daily (testing Ministry of the Interior and Safety, National Tax Service, Ministry of Government Legislation services)","sector":"National government resident-facing AI chat","geography":"South Korea","publishedAt":"2026-10-26","publicationDate":"2026-10-26","eventDate":"2026-10-26","sourceName":"Seoul Economic Daily","sourceLabel":"Seoul Economic Daily (English edition, AI-translated from Korean)","sourceUrl":"https://en.sedaily.com/technology/2026/10/26/government-ai-services-give-wrong-answers-to-simple","evidenceClass":"independent-reporting","outcomeClass":"cautionary","topics":["knowledge-work","governance-procurement","operating-model"],"finding":"Reporters queried three government AI services on the same day. AI Government24 said sole proprietors must file interim VAT returns, a requirement the article says was abolished in 2012, while the tax agency's Hometax chatbot said they are not required to. AI Government24 labeled that wrong answer and a wrong income-tax filing period as \"high confidence.\" On a pet-death question, two services pointed to different procedures. The interior ministry said it would standardize and structure government-wide civil service data to reduce wrong answers.","sledRelevance":"Relevant to U.S. 311 and digital front-door assistants: when several agencies run separate bots over separate content, residents can get contradictory answers, and self-reported confidence labels can mislead. Transferability is limited because these are national services in a different legal system, but the failure mode (stale source content, no cross-agency answer consistency) applies to multi-department city and county portals.","evidence":"Journalistic spot test of a handful of queries on one day; no sample size, error rate or systematic method. The article names the specific wrong answers and the correct rules for comparison. The English text is an AI translation of Korean reporting.","architectureImplications":"Separate bots over separate knowledge bases create inconsistent answers; a shared, current, structured content layer (the ministry's stated fix) and retrieval with source dating matter more than the model.","governanceImplications":"Agencies running multiple resident-facing assistants need cross-agency answer ownership, content expiry rules and a process for retiring superseded rules from source material.","securityPrivacyImplications":"No privacy incident reported. The risk is reliance harm: residents acting on wrong tax guidance.","caveats":"Small, unsystematic reporter test; AI-translated article; national rather than state or local government; the services' own accuracy statistics were not reported.","streamIds":["citizen-engagement"],"roles":{"sales":"Accuracy, not channel count, is the buying risk. This test shows a government bot confidently citing a rule abolished years ago and contradicting a sister agency's bot. For SLED buyers with several department chatbots, the conversation is about a shared, owned knowledge layer and answer testing before launch, not another assistant. Position bounded content-governance and evaluation work; do not promise error-free answers.","engineering":"Treat this as a retrieval and content problem. Conflicting answers came from separate services over separate data, and the interior ministry's own fix is to standardize and structure source data. Engineering should look at single-source content, effective-date metadata on rules, retrieval filters that exclude superseded material, and regression question sets run across every assistant. Confidence labels need calibration evidence or should be removed.","delivery":"Delivery should plan for ongoing content operations rather than a one-time launch: owners for each policy area, review cycles that retire outdated rules, a cross-assistant regression test run on every content change, and a resident path to report wrong answers. Acceptance should be measured on a fixed test set with known correct answers. This is interpretation drawn from a small reporter test."},"roleDetails":{"sales":{"customerProblem":"Residents get confident, outdated or contradictory answers when multiple government assistants draw on unmanaged content.","stakeholders":["Chief digital or resident-experience officer","311 or contact-center director","Department content owners","Legal or policy counsel"],"discoveryQuestions":["How many resident-facing assistants does the jurisdiction run, and do they share a knowledge source?","Who is accountable for retiring superseded rules from bot content?","Do any assistants display confidence scores, and has anyone checked their calibration?"],"valueHypothesis":"A shared, dated knowledge layer and cross-assistant testing could reduce contradictory answers; size of benefit is unmeasured in this source.","potentialEngagement":"A short answer-consistency audit across existing assistants using a fixed set of resident questions with verified answers.","unsupportedClaims":["Any error rate for AI Government24 or Hometax","That U.S. local assistants have the same defects","That a single knowledge base removes all wrong answers"]},"engineering":{"fit":"Applies to any multi-department chatbot, voicebot or portal search built on retrieval over agency content.","architectureIntegration":"Centralize or federate authoritative content with effective and expiry dates; filter retrieval on validity; run shared regression tests across all assistants.","prerequisites":["Inventory of assistants and their content sources","Content owners per topic","Maintained ground-truth question set"],"constraints":["Legal rules differ by jurisdiction and change often","Vendor platforms may not expose retrieval or confidence internals"],"security":"Main risk is reliance harm from wrong guidance; log answers and sources so errors can be traced and corrected.","proofOfValue":["Run 50+ fixed questions with verified answers across every assistant and measure disagreement","Check whether displayed confidence correlates with correctness","Seed a superseded rule and confirm it is not retrieved"]},"delivery":{"work":"Stand up content governance and cross-assistant answer testing for resident-facing AI.","dependencies":["Department content owners","Vendor access to content and logs","Ground-truth question set"],"ownership":"Resident-experience or 311 office, with department content owners accountable for accuracy.","skillsAdoption":"Content staff need training on writing dated, structured source content and reviewing bot logs.","governanceCheckpoints":["Pre-launch consistency test","Quarterly content expiry review","Review of any displayed confidence indicator"],"acceptanceCriteria":["Proposed: zero contradictory answers on the fixed test set before launch","Proposed: superseded rules removed within a set number of days of change"],"risks":["Content drift after launch","Residents acting on wrong answers","Confidence labels creating false trust"]}},"retrievedAt":"2026-10-30T03:09:17Z","enrichedAt":"2026-10-30T03:17:06Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Confidence labels shown to residents can mislead people least able to verify answers; staff still handle the fallout of wrong guidance.","procurementImplications":"Contracts for resident-facing assistants should require content-freshness controls, answer-consistency testing and calibrated (or removed) confidence displays.","operatingModelImplications":"Content governance is the operating bottleneck: someone must own retiring obsolete rules across departments.","sourceVerification":{"openedUrl":"https://en.sedaily.com/technology/2026/10/26/government-ai-services-give-wrong-answers-to-simple","referenceExcerpt":"More troubling is that the AI rated its own incorrect answer as highly reliable.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}