{"resourceId":"racovan-argus-llm-assisted-cs2-integrity-20260928","versions":[{"version":"external-2aecc9ebf44bea3ca61198d763c821d17b3c60ee9b34afcb55a5ef38713481ff","resource":{"id":"racovan-argus-llm-assisted-cs2-integrity-20260928","title":"Argus CS2 study: 45.7% Spring 2026 homework patterns flagged as LLM-assisted; exam correlations worsen","organization":"Purdue University (Racovan, Rawat, May, Turkstra)","sector":"Higher education computer science / academic integrity","geography":"United States (Purdue University large-enrollment CS2)","publishedAt":"Submitted to arXiv 28 September 2026 (cs.CY)","publicationDate":"2026-09-28","eventDate":null,"sourceName":"arXiv","sourceLabel":"Academic preprint (cs.CY)","sourceUrl":"https://arxiv.org/pdf/2609.36073","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["developers-agents","governance-procurement","data-security","operating-model","knowledge-work"],"finding":"Argus combines behavioral and stylistic indicators to surface possible LLM-assisted C programming homework across six Spring CS2 offerings. In Spring 2026, 267 of 584 students who completed the first midterm (45.7%) were flagged after manual review. Average H-scores were negatively correlated with performance on in-person closed-note exams, with correlations strengthening in more recent offerings. Authors stress that erroneous accusations carry high stakes and that any automated detection system must include structured human review; H-scores are a behavioral proxy, not verified proof of AI use.","sledRelevance":"Critical scrutiny companion to GenAI tutoring/adoption expansions: large-course homework assistance patterns are common, associate with weaker proctored performance, and demand due-process human review—not automatic discipline dashboards. Student-success leaders should pair AI help tools with assessment redesign and integrity processes that protect students from false positives.","evidence":"Longitudinal static-analysis of anonymized CS2 homework across six Spring offerings with the same instructor; Spring 2026 deployment with manual review of flags; correlations of H-score to midterm/final exam percentages. Observational; preprint.","architectureImplications":"If deploying integrity analytics, separate detection scores from disciplinary workflows; require human-review queues, explainable features, and audit logs. Prefer assessment designs (proctored/written) that measure unaided competence alongside homework help tools.","governanceImplications":"Policy must forbid sole-reliance on automated LLM-use flags for misconduct findings; publish review SLAs and student appeal paths. Align AI-tutor permissions with graded-work rules so help-seeking design does not silently become an integrity trap.","securityPrivacyImplications":"Homework and behavioral traces are sensitive education records. Minimize retention of full code in detector pipelines; restrict access to instructors/conduct officers under school-official rules; guard against biased false positives in large courses.","caveats":"Single course/instructor; proxy labels; false-positive risk acknowledged; correlations ≠ proven causation of learning loss; CS2 C-programming context; preprint.","streamIds":["student-success"],"roles":{"sales":"For provosts and CS deans scaling GenAI help while worried about integrity, Argus supplies measured caution: nearly half of a Spring 2026 CS2 midterm-completing cohort showed LLM-assisted homework patterns after human review, with worsening negative links to proctored exams. Sell assessment redesign plus human-reviewed analytics—not an auto-accuse product. Bound claims to Purdue CS2 observational evidence and refuse guaranteed detection accuracy or misconduct conviction rates.","engineering":"Engineering should treat detectors as decision-support services with mandatory human-review queues, calibrated thresholds, explainable features, and audit logs—never auto-write to conduct systems. Keep disciplinary actions write-gated behind instructor or integrity-officer decisions. Constraints include proxy labels rather than ground-truth AI use and single-course evidence requiring local revalidation. Proof-of-value is reviewer precision under load and reduced false-positive escalations, not raw flag volume. Minimize retention of full student code in detector pipelines.","delivery":"Delivery should stand up instructor review playbooks, clear student-facing AI homework rules, and parallel assessment redesign that preserves proctored measures of unaided competence. Ownership sits with the course instructor of record, with the academic integrity office handling escalations after human review. Communicate that flags are risk signals, not verdicts. Risks include mass false accusations, chilling legitimate help-seeking, and confusing tutoring bots with cheating detectors in student communications."},"roleDetails":{"sales":{"customerProblem":"Large CS courses see widespread suspected LLM homework help while needing fair integrity processes and unaided learning measures.","stakeholders":["CS department","academic integrity office","provost","student success","teaching & learning"],"discoveryQuestions":["What share of graded work is take-home vs proctored?","Who reviews integrity analytics before student contact?","How do AI tutor policies interact with homework rules?"],"valueHypothesis":"Human-reviewed integrity analytics plus assessment redesign can surface learning risk without automatic misconduct findings.","potentialEngagement":"Shadow-mode Argus-like review for one CS2-style course with documented human-review SLA and exam-outcome monitoring.","unsupportedClaims":["Verified individual AI use from H-score alone","Causal proof H-score causes exam decline","Cross-course detector portability without revalidation"]},"engineering":{"fit":"Large programming courses with versioned homework submissions and separate proctored exams.","architectureIntegration":"Submission ingest → feature/H-score service → instructor review queue → SIS/conduct only after human decision; audit trail.","prerequisites":["historical homework corpus","instructor review capacity","clear academic-policy mapping"],"constraints":["proxy not ground truth","false positives","single-instructor evidence base"],"security":"Code and scores as education records; least-privilege reviewer access; no auto-sanction API; retention limits.","proofOfValue":["Manual-confirm rate among flags","Time-to-review SLA","Correlation monitoring without auto-discipline"]},"delivery":{"work":"Configure detector, train reviewers, communicate policy to students, redesign assessments where needed.","dependencies":["faculty time","integrity office procedures","student appeal channels"],"ownership":"Course instructor of record with academic integrity office for escalations.","skillsAdoption":"Reviewers interpret behavioral proxies; students taught legitimate vs prohibited AI help.","governanceCheckpoints":["Policy sign-off before flags contact students","weekly false-positive review","end-term process audit"],"acceptanceCriteria":["No misconduct finding without human review","Published student-facing AI homework rules","Proctored measures retained for competence"],"risks":["Mass accusations","Chilling effect on help-seeking","Over-trust in H-score"]}},"retrievedAt":"2026-09-30T03:28:26Z","enrichedAt":"2026-09-30T03:31:28Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"False-positive integrity flags can disproportionately harm students without advocacy capital; human review and clear appeal paths are equity requirements. Faculty workload for review is a real staffing cost of detection.","procurementImplications":"Do not buy integrity products that auto-accuse. Require human-in-the-loop workflows, explainability, and independent evaluation on local corpora. Separate tutoring-tool procurement from detection-tool procurement with aligned policies.","operatingModelImplications":"Operate GenAI student help and integrity analytics as linked but gated systems: help tools with progressive assistance; detectors with review queues; proctored assessments as competence anchors.","sourceVerification":{"openedUrl":"https://arxiv.org/pdf/2609.36073","referenceExcerpt":"any automated detection system must be accompanied by a structured process for human review","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}