Higher Education & ResearchIssue 125 records
Student Success
Five newly archived higher-education sources after 2026-09-16: a Flinders Cogniti GenAI academic-agent pilot finds perceived learning/success gains with flat engagement/wellbeing; a June 2026 Frontiers meta-analysis reports large but bias-prone intellectual effects and smaller social–emotional effects across 33 studies; a Massey doubly robust evaluation of AI-guided proactive support estimates fewer failures and ~8.8-point higher grades with weaker completion-speed effects; Qian’s analysis of 50 innovative U.S. universities maps syllabus-led GenAI governance, privacy guardrails, and detector caution; and Wright argues integrity policies must separate assistive transcription/OCR from generative authorship. Two patterns stress evaluation-over-access and governance precision for assistive vs generative tools. U.S. CC persistence RCTs, disability causal trials, and TCO remain gaps.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 5 academic research
- Outcomes
- 3 mixed1 emerging1 cautionary
- Source freshness
- 3 older, newly relevant2 recent
- Research completed
- 2026-09-18
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Student-success AI value depends on design plus evaluation—not access or CSAT alone
Cogniti’s agent pilot lifted only some perceptions; the HE GenAI meta-analysis shows large but biased/heterogeneous intellectual effects; Massey’s AI-guided support evaluation finds failure/grade gains with weaker completion-speed effects. Together they argue campuses must instrument differentiated outcomes and causal-ish evaluation when scaling agents or early-alert.
Operating questionWhich pre-registered outcomes (perceptions, intellectual learning, failures/grades, completion speed) will govern go/no-go for each student-success AI use case?
Supporting evidenceFlinders University / Technology, Knowledge and Learning (Springer)Chuxiong Normal University / Sichuan University of Science & Engineering / Frontiers in PsychologyMassey University (Auckland) / arXiv preprint
Integrity and privacy governance need technical precision and vetted-tool ecosystems
Qian’s 50-university analysis converges on syllabus-led rules, disclosure, privacy guardrails, detector caution, and literacy; Wright argues policies must separate non-generative transcription/OCR from generative authorship to avoid disability harm. Student-success AI programs should adopt both pedagogy-first ecosystems and tool-class distinctions.
Operating questionDo our student-success AI policies define generation versus assistive format conversion and route tools through vetted, data-class-aware platforms?
Supporting evidenceLouisiana State University / International Journal for Educational IntegrityUniversity of Exeter / International Journal for Educational Integrity
Evidence ledger
Full record · every source keeps its link and limitations
Cogniti GenAI academic-agent pilot: perceived learning/success up; engagement and wellbeing flat
Pilot study of course-embedded Cogniti GenAI academic agents for assessment preparation across four undergraduate courses (education, psychology, social work, sport). N=139 survey completers (ages 18–55) pre/post agent deployment. Mann–Whitney U tests: significant improvements in perceived student learning and perceived success (both p=0.04; w≈1642 and 1639.5), with no significant change in engagement (p=0.30) or wellbeing (p=0.59). Qualitative themes valued 24/7 context-specific support and confidence; students also stressed retaining human academic interaction. Authors conclude agents may enhance targeted aspects of experience when carefully integrated to complement—not replace—human-centred learning.
Why it matters, evidence and limitations
- Why it matters
- Newly relevant after 2026-09-16 archived hybrid/reflective feedback and RAG-tutor nulls: this unarchived 2026 OA pilot supplies HE evidence on personalized academic agents’ effects on student-success perceptions versus engagement/wellbeing. Why now: fills agent-product gap for advising/tutoring-adjacent on-demand support without overclaiming objective grades.
- Evidence and measured results
- Source findings: pre/post perception surveys on one assessment; Mann–Whitney and rank-biserial effect sizes; content analysis of open responses. Distinguishes perception gains (learning/success) from null engagement/wellbeing. Limitations: pilot/non-randomized; self-report only; single-assessment window; Australian one-university sample; response rates vary by course (e.g., psychology 9.6% of enrolled).
- Limitations and uncertainty
- Perception outcomes ≠ measured GPA/retention. Non-experimental design cannot support causal claims. Not a U.S. community-college advising RCT. Publication accepted late 2025 / ©2026; included as newly archived for this stream.
Meta-analysis: GenAI shows large but heterogeneous intellectual gains and smaller social–emotional effects in HE
Systematic review and random-effects meta-analysis of 33 experimental/quasi-experimental studies (3,394 participants; studies published 2021–2024; search through 26 Jan 2026). Generative AI showed a significant pooled positive effect on intellectual outcomes (Hedges’ g=1.096, 95% CI 0.087–2.104, p=0.033) and a smaller significant effect on social–emotional outcomes (g=0.301, 95% CI 0.048–0.553, p=0.020). Domain difference was not statistically significant. Publication bias detected for intellectual outcomes; social–emotional findings more stable in sensitivity analyses. Authors conclude GenAI is promising but evidence remains heterogeneous and methodologically uneven (PROSPERO CRD420261361622).
Why it matters, evidence and limitations
- Why it matters
- Newly relevant synthesis for student-success leaders after single-study mixed RCT nights (2026-09-15/16): provides an unarchived 2026 pooled view of intellectual vs social–emotional effects with explicit bias/heterogeneity warnings useful for portfolio evaluation.
- Evidence and measured results
- Source findings: Stata 18 random-effects models; moderator analyses (functional type, duration, knowledge domain) not significant at omnibus level; funnel/sensitivity for bias. Limitations: wide CI on intellectual g; detected publication bias; included studies to 2024 only; HE contexts heterogeneous; social–emotional construct breadth.
- Limitations and uncertainty
- Large intellectual point estimate sits in a very wide CI with publication bias—do not market g≈1.1 as expected local effect. Not U.S.-only; not advising/retention specific. Published June 2026 (before last run) but not previously archived for this stream.
Doubly robust evaluation: AI-guided proactive support cuts failures and lifts grades; completion speed weaker
Observational causal evaluation of an institutional AI learning-analytics intervention that produces daily probability-of-success scores used by a central student-support team for proactive outreach to at-risk undergraduates. Using blocked/calipered propensity-score matching and doubly robust estimation with bootstrap uncertainty, authors report the intervention reduced course failure rates and raised cumulative grades (treated students estimated ~8.8 grade points higher on average than matched controls; treated students failed about 0.78 fewer courses on average). Effects on speed of qualification completion were positive but statistically constrained. Activity Theory framing: AI risk scores mediate support work and stabilize trajectories, but institutional secondary contradictions limit translation into faster completion. Ethics approval Massey HE014–OM3 23/42.
Why it matters, evidence and limitations
- Why it matters
- Newly relevant after 2026-09-16 ACC Digital Twin reporting (no outcomes): this unarchived evaluation supplies Pass A measured effects for AI-guided proactive support—the advising/early-alert gap called out in prior editions—while Pass B notes non-RCT design and weak completion-speed effects.
- Evidence and measured results
- Source findings: PSM + doubly robust ATT-style estimates; highlights claim reduced failures and improved grades/pass rates; AI model performance metrics reported (e.g., F1/AUC on risk detection). Limitations: not a randomized trial; residual confounding possible despite sensitivity analysis; single NZ university; preprint; completion-speed effects constrained; immortal-time addressed via dynamic scores but design remains observational.
- Limitations and uncertainty
- December 2025 preprint (older than last run) included because AI-guided support outcome evidence was a stated archive gap and URL was check_sources-unknown. Observational identification ≠ RCT. Grade-point metric may be local scale. Transfer to U.S. community colleges requires re-estimation.
Fifty innovative U.S. universities: syllabus-led GenAI governance, privacy guardrails, detector caution
Qualitative document analysis of official websites for the 50 U.S. universities ranked most innovative by U.S. News identifies five convergent GenAI teaching/learning patterns: (1) instructor-led, syllabus-level governance; (2) disclosure and attribution expectations; (3) privacy-oriented data guardrails and vetted tools; (4) caution toward AI-detection systems; and (5) expanding AI literacy investment. Divergences remain in default permission stances, enterprise governance maturity, and tooling (campus-hosted vs consumer tools under risk controls). Support is distributed across teaching centers, libraries, IT/security, and integrity offices. Author argues integrity is shifting from punitive models toward pedagogy-first ecosystems combining expectations, assessment redesign, and equitable access to vetted tools.
Why it matters, evidence and limitations
- Why it matters
- Pass B governance update after 2026-09-16 Ng et al. 43-university privacy guideline study: this unarchived 2026 U.S.-innovative-50 analysis adds detector caution, syllabus-level defaults, and distributed support ecosystems—actionable for student-success integrity and advising-tool approvals.
- Evidence and measured results
- Source findings: inductive thematic analysis of institutional web guidance; convergence/divergence taxonomy; examples include FERPA-sensitive data classifications and vetted campus platforms (e.g., Michigan data-class limits cited). Limitations: document analysis ≠ practiced enforcement; U.S. News innovative sample skews research universities; websites change; no student-outcome measures.
- Limitations and uncertainty
- Accepted July 2026 (before last run) but not archived previously. Policy texts may lag practice. Innovative-university sample may not transfer to community colleges without localization.
Transcription ≠ generation: integrity policies must separate assistive AI format conversion from GenAI authorship
Conceptual/policy analysis arguing HE GenAI assessment policies often lack technical precision to distinguish generative AI (machine produces assessed intellectual content) from non-generative AI-powered tools—OCR, voice-to-text transcription, handwriting recognition—that perform clerical format conversion on student-authored content. Notes widespread faculty-facing GenAI guidelines (e.g., An et al. 2025 on top U.S. universities) and UK QAA/Russell Group guidance. Argues treating accessibility transcription as misconduct raises fairness concerns under equality law (UK Equality Act reasonable-adjustment duties; parallels to ADA/Australian DDA). Recommends explicit policy language and an evidential framework separating generation from transcription.
Why it matters, evidence and limitations
- Why it matters
- Pass B scrutiny complementary to 2026-09-16 Shepherd Against Prohibition: adds technical policy precision on assistive vs generative tools—critical as student-success units deploy voice/OCR supports alongside GenAI tutors. Unarchived; accepted Aug 2026.
- Evidence and measured results
- Source type: conceptual and policy analysis with interpretive arguments and proposed evidential framework—not an empirical outcomes study. Limitations: UK-weighted legal framing; tribunal outcomes not claimed; depends on cited secondary surveys of university guidelines; no measured false-positive rates in this paper.
- Limitations and uncertainty
- Not empirical; legal characterizations are interpretive. Transfer to U.S. requires counsel review of ADA/Section 504 interplay with AI integrity rules. Accepted Aug 2026 but not previously archived.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.