{"resourceId":"akcapinar-lms-early-ai-cheating-risk-20260929","versions":[{"version":"external-1cb4b0d000477b4708642c9a14f0ccc4d7320898f132a25ee20474ed4964278d","resource":{"id":"akcapinar-lms-early-ai-cheating-risk-20260929","title":"LMS traces from weeks 1–8 predict AI-assisted exam-cheating risk at 73.1% accuracy—for guidance, not discipline","organization":"Hacettepe University (Gökhan Akçapınar)","sector":"Higher education computing education / academic integrity analytics","geography":"Türkiye (public university; first-year Computer Education and Instructional Technology undergraduates)","publishedAt":"Submitted to arXiv 29 September 2026 (cs.CY); to be presented at ICETC 2026","publicationDate":"2026-09-29","eventDate":null,"sourceName":"arXiv","sourceLabel":"Academic preprint (cs.CY)","sourceUrl":"https://arxiv.org/pdf/2609.37280","evidenceClass":"academic-research","outcomeClass":"emerging","topics":["data-security","governance-procurement","developers-agents","operating-model"],"finding":"Among 52 first-year undergraduates in an Introduction to Programming course, 23 (44.2%) were labeled high-risk for AI-assisted cheating based on suspicious final-exam log behaviors (copy, focus-loss, right-click) in a proctored face-to-face exam. Using five features from 27 early-semester LMS/video candidates, Logistic Regression under LOOCV with fold-specific preprocessing achieved Accuracy 73.1% (Balanced Accuracy 0.718, AUC 0.760). Most stable predictors—course-module views, assignment submissions, and video-access days—were lower in the high-risk group. Authors frame predictions for timely academic guidance, not misconduct findings; labels are behavioral proxies, not verified AI use.","sledRelevance":"Offers HE student-success teams an early-warning pattern complementary to post-hoc homework detectors: thin early engagement may foreshadow integrity-risk behaviors, but only as outreach triggers with human judgment. Reinforces that analytics must stay outside automatic discipline pipelines.","evidence":"LOOCV classification on n=52 with exam-log-derived risk labels and early LMS/video features; feature selection via Mann-Whitney U per fold. Small single-course sample; proxy labels; preprint/conference paper.","architectureImplications":"If building early alerts, score LMS engagement features in week 8, route to advisor/instructor outreach queues, and hard-separate from conduct systems. Calibrate thresholds for high recall outreach vs high precision escalation.","governanceImplications":"Policy must state that LMS risk scores are guidance signals only; prohibit sole use for misconduct findings; require human review, student-facing AI-use guidance, and appeals. Local revalidation required before transfer from Türkiye programming context.","securityPrivacyImplications":"LMS and exam-behavior traces are sensitive education records. Minimize feature sets, restrict reviewer access, avoid storing raw keystroke/focus logs longer than needed, and prevent vendor secondary use.","caveats":"n=52; single course/country; high-risk label from exam-log proxies not confirmed AI use; 73.1% accuracy leaves substantial error; LOOCV optimism risk; preprint.","streamIds":["student-success"],"roles":{"sales":"For student-success and integrity leaders seeking earlier intervention than end-of-term detectors, this study shows week-1–8 LMS traces can classify AI-assisted exam-risk labels at 73.1 percent accuracy in one programming course—while authors insist on guidance, not discipline. Sell a shadow-mode early-outreach pilot with human advisors, not an auto-accuse product. Bound claims to n=52 proxy-labeled evidence and refuse guaranteed misconduct detection or portable accuracy.","engineering":"Engineering should compute early engagement features such as module views, assignment submissions, and video-access days, emit risk scores to an advisor queue, and block any API path into conduct case systems. Support threshold presets for outreach versus review. Constraints include small-N proxy labels and cross-border transfer limits requiring local revalidation. Proof-of-value is advisor-actioned outreach with documented false-positive handling—not raw accuracy. Keep exam-proctoring raw logs tightly access-controlled under school-official rules.","delivery":"Delivery should run a non-punitive week-8 outreach protocol: instructors or advisors contact flagged students with AI-use guidance and academic support, logging outcomes without opening misconduct cases from the score alone. Ownership sits with the course instructor and advising lead; the integrity office engages only after separate human-verified evidence. Acceptance means written guidance-only policy and local revalidation metrics. Risks include stigmatizing low-engagement students and exporting a Türkiye programming model without recalibration."},"roleDetails":{"sales":{"customerProblem":"Integrity risk surfaces too late; campuses want early signals without automatic misconduct pipelines.","stakeholders":["student success","academic integrity office","CS/education faculty","advising","registrar/LMS owners"],"discoveryQuestions":["Who receives early risk alerts?","What outreach happens before any allegation?","How will we revalidate on local LMS data?"],"valueHypothesis":"Early LMS engagement alerts can trigger timely academic guidance when human-reviewed and policy-gated away from discipline.","potentialEngagement":"One-course shadow-mode early-warning pilot with advisor outreach scripts and accuracy audit.","unsupportedClaims":["Verified individual AI cheating from LMS score","73.1% portable accuracy","Replacement for assessment redesign"]},"engineering":{"fit":"LMS-integrated early-alert services for programming or online-assessed HE courses.","architectureIntegration":"LMS/video feature job → score service → advisor queue UI → no write path to conduct; audit log of views/actions.","prerequisites":["LMS API access","advisor staffing","local labeled validation set or proxy protocol"],"constraints":["proxy labels","small-N source","jurisdictional transfer"],"security":"Education-record minimization; role-based reviewer access; retention limits on proctoring artifacts.","proofOfValue":["Outreach completion rate","False-positive review notes","No misconduct cases opened from score alone"]},"delivery":{"work":"Define guidance scripts, train advisors, run week-8 scoring, document non-punitive contacts, revalidate features.","dependencies":["LMS data steward","course faculty","advising capacity"],"ownership":"Course instructor with student-success advising; integrity office only on independent evidence.","skillsAdoption":"Advisors deliver AI-use guidance and support without accusatory framing.","governanceCheckpoints":["Policy sign-off guidance-only","threshold review","end-term disparity audit"],"acceptanceCriteria":["Published guidance-only rule","Human contact before any allegation","Local performance memo"],"risks":["Stigma of low engagement","Threshold gaming","Cross-context model failure"]}},"retrievedAt":"2026-10-01T03:23:54Z","enrichedAt":"2026-10-01T03:26:30Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Low early LMS engagement may reflect work hours, disability, or access barriers—not cheating intent. Outreach must offer support resources and avoid deficit labeling; advisor workload is a real staffing cost.","procurementImplications":"Prefer early-alert products with human-outreach workflows, explainable features, and contractual bans on automatic disciplinary actions. Require local revalidation clauses.","operatingModelImplications":"Operate early AI-integrity risk scores as student-success outreach, parallel to—but gated from—conduct processes; pair with assessment redesign and clear AI-use guidance.","sourceVerification":{"openedUrl":"https://arxiv.org/pdf/2609.37280","referenceExcerpt":"These predictions are intended to support timely academic guidance, not to establish misconduct or initiate disciplinary action","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}