{"resourceId":"poudyal-genai-agency-accountability-20260923","versions":[{"version":"external-4e1dd32b95511caa6de258d7b5650dd4a8ecc30780c39f7ae16e80075f77b431","resource":{"id":"poudyal-genai-agency-accountability-20260923","title":"Discourse analysis: HE GenAI research grants AI agency but never makes AI responsible","organization":"Charles Darwin University","sector":"Higher education / academic research","geography":"International English-language HE research corpus (transferable as discourse evidence)","publishedAt":"2026-09-23","publicationDate":"2026-09-23","eventDate":null,"sourceName":"arXiv","sourceLabel":"Preprint (arXiv:2609.27184)","sourceUrl":"https://arxiv.org/pdf/2609.27184","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["governance-procurement","operating-model","knowledge-work","accessibility-workforce"],"finding":"Across 366 HE GenAI titles/abstracts (91,405 tokens), AI is the most frequent named actor (2,050 mentions) with predominantly action predicates, while students are cast as knowers/judges/verifiers; no sentence makes AI explicitly responsible. Responsibility lands on educators, institutions, and policy or disappears in passives and nominalisations—even when texts invoke 'responsible AI.'","sledRelevance":"Warns U.S. SLED student-success leaders that policy and vendor language celebrating AI agency can outrun named human accountability for tutoring, advising, and assessment decisions.","evidence":"Corpus-assisted discourse analysis with frequency, collocation, actor–predicate, obligation-clause, and passive/nominalisation heuristics on Scopus/WoS/Springer/IEEE-derived records after deduplication.","architectureImplications":"Design logs and UI copy that attribute actions to named human roles (instructor, advisor, student) even when models generate drafts; avoid agent-washing in product claims.","governanceImplications":"Rewrite 'responsible AI' policies to state who authorizes, who verifies, and who answers for consequences of GenAI-shaped student work and support.","securityPrivacyImplications":"Accountability gaps compound privacy risk when no owner is named for data entered into tutors/advising bots.","caveats":"Titles/abstracts only; incomplete 2026; not an outcomes evaluation.","streamIds":["student-success"],"roles":{"sales":"For provosts and academic-integrity buyers, this discourse study shows HE research already treats GenAI as an active doer while never calling AI responsible—so 'responsible AI' marketing without named human owners is a red flag. Position discovery around RACI for tutoring/advising GenAI and obligation clauses in policy—not anthropomorphic agent demos. Bounded engagement is a policy/language audit workshop, not a claim that discourse fixes learning outcomes.","engineering":"Engineering takeaway: instrument provenance that records which human approved AI-generated feedback, advice, or assessment artifacts; avoid product copy that assigns epistemic authority to the model. Prerequisites include role-based approval workflows and auditable logs. Constraints: external evidence is discourse, not a deployment RCT. Security: clear data stewards when no 'AI owner' exists. Proof of value: policy and UI texts that name accountable roles for each GenAI action class.","delivery":"Delivery work: inventory campus GenAI student-success tools, map who acts, who knows, and who answers for each workflow, and revise syllabus or policy language that uses agent metaphors without named owners. Dependencies include academic affairs, the integrity office, and IT governance. Ownership sits with a provost delegate partnered with teaching-and-learning. Skills: policy drafting and faculty communication. Governance checkpoints: obligation-clause review before launch. Acceptance: published RACI per tool; risks include performative 'responsible AI' stickers without enforcement."},"roleDetails":{"sales":{"customerProblem":"Campuses buy 'agentic' student-success AI while policies never name who answers for errors.","stakeholders":["provost/academic affairs","integrity office","IT governance","student success"],"discoveryQuestions":["Who is answerable when a tutor over-helps on graded work?","Do policies define responsible use with named roles?","Where does vendor language assign agency to the model?"],"valueHypothesis":"Bounded value: clarify human accountability language before scaling agentic tutoring/advising.","potentialEngagement":"2–3 week discourse/policy audit of GenAI student-success tools—not an outcomes guarantee.","unsupportedClaims":["Discourse analysis proves learning harm","AI systems can be made legally responsible","Guaranteed integrity improvement from wording changes alone"]},"engineering":{"fit":"Fits as governance/provenance layer across tutoring and advising GenAI services.","architectureIntegration":"Action log with human approver IDs; policy engine for obligation clauses; UI attribution strings.","prerequisites":["Role directory","Audit logging","Policy CMS"],"constraints":["Evidence is corpus discourse","Title/abstract bias","No deployment metrics in source"],"security":"Assign data stewards; prevent unowned model access to education records.","proofOfValue":["Named owner per GenAI action class","Zero 'AI is responsible' product claims in campus UI","Obligation clauses in published policy"]},"delivery":{"work":"Run corpus-informed language audit; rewrite policies; train staff on attribution.","dependencies":["Academic affairs","Legal/integrity","Communications"],"ownership":"Academic affairs with IT governance co-owner.","skillsAdoption":"Workshop faculty/advisors on who acts vs who answers.","governanceCheckpoints":["Policy obligation-clause review","Vendor contract accountability schedule","Syllabus template update"],"acceptanceCriteria":["RACI published for each GenAI student-success tool","Sample syllabi name human verification duties"],"risks":["Symbolic compliance","Vendor pushback on logs","Over-reading abstracts as campus practice"]}},"retrievedAt":"2026-09-29T03:18:41Z","enrichedAt":"2026-09-29T03:21:30Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"When accountability is unnamed, disabled students relying on GenAI assistance face unclear appeal paths—policies should state human review rights explicitly.","procurementImplications":"Require vendors to identify customer-side accountable roles and logging for AI-generated student-facing outputs; reject agent claims without human override paths.","operatingModelImplications":"Separate functional automation metrics from accountability assignments; escalate unresolved 'who answers' gaps before scaling bots.","sourceVerification":{"responseId":"curl-arxiv-2609-27184","openedUrl":"https://arxiv.org/pdf/2609.27184","referenceExcerpt":"The findings reveal a persistent gap between giving AI functional agency and assigning normative accountability.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}