{"resourceId":"augie-mirror-ask-augustana-chatbot-test-20260925","versions":[{"version":"external-af3eb2847c2ca43b1e53e35f97b2b249d3ccbf723be5c34b25ac309d03f4976b","resource":{"id":"augie-mirror-ask-augustana-chatbot-test-20260925","title":"Augie Mirror test of Augustana's Gemini-based 'Ask Augustana' chatbot finds stale answers, prompt disclosure and scope drift","organization":"Augustana University (South Dakota) / Zion Cloud Solutions (via The Augie Mirror)","sector":"Private higher education / website and student-services AI","geography":"United States (South Dakota)","publishedAt":"September 25, 2026","publicationDate":"2026-09-25","eventDate":"2026-08-15","sourceName":"The Augie Mirror (student newspaper)","sourceLabel":"Independent student-journalist testing of an institutional AI tool","sourceUrl":"https://www.augiemirror.news/2026/09/25/asking-ask-augustana-putting-the-universitys-new-chatbot-to-the-test/","evidenceClass":"independent-reporting","outcomeClass":"cautionary","topics":["knowledge-work","data-security","governance-procurement","operating-model"],"finding":"Over several weeks after the Aug. 15 launch, Mirror reporters tested 'Ask Augustana', a custom Google Gemini assistant built with Zion Cloud Solutions to unify the university's websites. Leaders said the grounded agent 'doesn't hallucinate.' Testing found it gave a student the opposite of Financial Aid's guidance on a meal-plan scholarship question, described a DEI office dissolved in 2025 and a discontinued major as current, repeated a former professor as department chair when challenged, and said campus safety officers are unarmed (faithful to an outdated page). Rephrased prompts produced near-complete disclosure of its rules and tool name, and it answered off-scope math and Klingon requests despite an English-only patch.","sledRelevance":"Direct campus-operations signal: public-facing admissions, aid and student-life chatbots inherit every stale page on institutional sites, so content governance and web lifecycle become AI operating prerequisites. The university plans an advising agent next, raising the stakes.","evidence":"Opened the Mirror article (Sept. 25, 2026) describing specific test prompts, screenshots and responses, plus on-record statements from the VP for Innovation and the Zion Cloud developer.","architectureImplications":"Retrieval-grounded design over multiple domains (augie.edu, library, athletics) did not prevent stale-source answers or missed content on a separate catalog domain (smartcatalogiq); source freshness metadata and archived-page exclusion must be engineered, not assumed.","governanceImplications":"Leadership framed errors as 'garbage in, garbage out', which places accountability on web content owners; institutions need named owners for each indexed source, takedown of retired pages, and a correction loop before expanding to advising.","securityPrivacyImplications":"Prompt rephrasing extracted system rules and internal tool names; scope and language restrictions were bypassed. No sensitive data exposure was reported, but the same weaknesses would matter for an advising agent that touches student records.","caveats":"Student-newspaper testing is non-systematic, with no stated sample size or error rate; some answers changed during testing as developers patched issues. The vendor and university dispute that the model itself hallucinates.","streamIds":["campus-operations"],"roles":{"sales":"Sales framing: a chatbot is only as accurate as the content estate behind it. Augustana's launch shows leaders promising 'no hallucination' while students got contradictory aid guidance and retired offices. Lead with a content-readiness conversation, not a bot demo: who owns each source page, how retired content is removed, and how errors reach staff. A bounded content audit plus red-team test is credible; promising accuracy improvements or reduced call volume from this evidence is not.","engineering":"Engineering framing: grounding alone did not stop wrong answers. Fit is any retrieval assistant indexing sprawling .edu sites. Architecture should add source freshness metadata, exclusion of archived pages, coverage of separate catalog domains, and visible citations. Security testing must include indirect prompt disclosure and scope bypass, both of which the Mirror achieved with simple rephrasing. Proof of value should be a fixed question set scored against current policy owners' answers, rerun after every content change.","delivery":"Delivery framing: the work is content operations as much as AI operations. Before an advising agent, Augustana-style programs need page ownership, retirement workflows, and a feedback path from students to content owners. Delivery should stand up a regression test suite of high-stakes questions (aid, safety, programs) with owner sign-off, track error-correction time, and gate expansion on measured accuracy. Risks include student harm from aid misinformation and reputational damage from public testing."},"roleDetails":{"sales":{"customerProblem":"Public-facing campus chatbot gives confident but outdated or contradictory answers on aid, programs and safety because it retrieves from stale institutional pages.","stakeholders":["VP for innovation / CIO","Web and marketing content owners","Financial aid director","Registrar and academic affairs","Campus safety communications"],"discoveryQuestions":["Who owns each domain and page the chatbot indexes, and how are retired pages removed?","How are wrong answers reported and how fast are they corrected?","Has the bot been tested on high-stakes aid and safety questions against current policy?","Is the catalog on a separate domain included in retrieval?"],"valueHypothesis":"Interpretation: a content-readiness audit and red-team test could reduce high-stakes misinformation before the institution expands to advising agents; size of benefit is unknown.","potentialEngagement":"A bounded 4-6 week content-estate audit and chatbot accuracy and prompt-security test, ending with a prioritized remediation list and regression question set.","unsupportedClaims":["The evidence does not show any measured error rate or call deflection.","It cannot support claims that grounding eliminates hallucination or wrong answers.","It does not show student harm beyond individual reported cases."]},"engineering":{"fit":"Retrieval-grounded assistants over multi-domain .edu web estates used for admissions, aid and student-life questions.","architectureIntegration":"Add freshness and status metadata to indexed pages, exclude archived content, include catalog domains, enforce citation display, and log answers for review.","prerequisites":["Inventory of indexed domains and owners","Page retirement and archive policy","Current authoritative answers for high-stakes questions","Answer logging with privacy review"],"constraints":["Separate vendor-hosted domains may not be crawled","Public bot cannot verify student identity or individual eligibility","Vendor model and prompt layer controlled by third party"],"security":"Test for system-prompt and tool-name disclosure via rephrasing, scope bypass, and instruction override; treat future advising-agent data access as a separate, higher-risk authorization.","proofOfValue":["Score a fixed set of 50+ aid, program and safety questions against owner-approved answers","Rerun the set after content changes to detect regressions","Run a prompt-disclosure and scope-bypass test script and record pass/fail"]},"delivery":{"work":"Stand up content governance, a high-stakes regression test suite and an error-correction workflow for the public chatbot before adding advising capabilities.","dependencies":["Web content owners across offices","Vendor (Zion Cloud-type) change process","Financial aid and registrar policy owners"],"ownership":"A named service owner in the innovation or IT office, with each content domain owned by its operating office.","skillsAdoption":"Content owners need training on how their pages feed the bot and how to retire or flag outdated material; staff need a reporting path for bad answers.","governanceCheckpoints":["Content-owner sign-off on high-stakes answers before launch of new scopes","Prompt-security test pass before any expansion","Advising-agent data access review before student-record integration"],"acceptanceCriteria":["Proposed: at least 95% correct on the owner-approved high-stakes question set","Proposed: reported errors corrected within 5 business days","Proposed: no system-prompt disclosure in the standard test script"],"risks":["Students acting on contradictory aid guidance","Public reputational damage from student testing","Scope creep into advising before content is fixed"]}},"retrievedAt":"2026-10-04T21:52:38-05:00","enrichedAt":"2026-10-04T21:52:38-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"The bot complied with an 'accessibility accommodation' request to speak like a pirate, showing accommodation framing can be used to override behavior; staff still carry correction work, so the tool shifts rather than removes front-office labor.","procurementImplications":"Contracts with AI integrators should require accuracy testing on institution-defined question sets, prompt-security testing, content-coverage disclosure and a correction SLA, rather than accepting 'grounded, does not hallucinate' claims.","operatingModelImplications":"Moves website stewardship into the AI service chain: content retirement and ownership become operational controls for any student-facing agent.","sourceVerification":{"openedUrl":"https://www.augiemirror.news/2026/09/25/asking-ask-augustana-putting-the-universitys-new-chatbot-to-the-test/","referenceExcerpt":"the chatbot is, so far, only as reliable as the web pages behind it.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}