Higher Education & ResearchIssue 525 records
Student Success
For 2026-10-29, the student-success signal is the gap between campus AI access and control over how that access shapes learning. NYU contract faculty asked administrators to disable Gemini's ability to edit students' Google Docs, because the edits are recorded as the student's own (Washington Square News, 28 Oct). A WashU psychology professor replaced writing assignments with in-class exams after colleagues could not tell AI writing from student writing (Student Life, 29 Oct). Notre Dame published summit findings that students link harmful AI use to workload and grade pressure and want assignment-specific guidance (28 Oct). Two new preprints add detail. A survey of 891 college students ties GenAI use to critical thinking only when students rework the AI's output. A Boston University benchmark found 13 LLMs differed widely in defending correct statics work under student pushback (both 28 Oct). Operating takeaway (interpretation): pair AI access with faculty governance input, configurable defaults, assessment-redesign support, reworking-focused literacy and tutor reliability checks. Limits: no government evaluation or controlled institutional outcome study dated 28 to 29 Oct, and the preprints are unreviewed.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 3 academic research2 independent reporting
- Outcomes
- 3 cautionary1 emerging1 mixed
- Source freshness
- 5 new this fortnight
- Research completed
- 2026-10-30
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Campus-wide AI tools are arriving faster than faculty control and assessment support
At NYU, a contract faculty council asked administrators to disable Gemini's ability to create and edit students' Google Docs, because the edits show as the student's own in revision history and work against course AI policies. NYU's AI Leadership Council will review the request (Washington Square News, 28 Oct). At WashU, where the +AI initiative supplies enterprise tools and a mandatory Canvas module, an AI-using psychology professor replaced writing with in-class exams after colleagues could not tell AI writing from student writing, and said he gets no central guidance on assessment. A creative writing director said he ignores +AI emails (Student Life, 29 Oct). Interpretation: enterprise AI rollouts need configurable defaults, faculty governance input and practical assessment support, or instructors retreat to narrower assessment formats. Limits: two news reports of a few voices, with no outcome data.
Operating questionBefore enabling campus AI tools for students, have we documented which embedded features are on by default, given faculty governance a say, and funded assessment-redesign support so instructors are not pushed into narrower formats?
Supporting evidenceNew York University (Full-Time Continuing Contract Faculty Senators Council; AI Leadership Council)Washington University in St. Louis (+AI initiative; Psychological & Brain Sciences; English)
How students use AI, and how reliably the AI responds, matters more than whether they have access
A preprint survey of 891 US and Turkish college students (28 Oct) found that GenAI experience related to critical thinking only when students divided work with the tool and reworked its output, not just from familiarity or confidence. A Boston University benchmark (28 Oct) found models differed widely in whether they held onto correct statics solutions when a student pushed back: they rejected work they had solved themselves in 1.6% to 50% of cases. Notre Dame's summit findings (28 Oct) show students and faculty want AI to complement rather than replace student thinking, through scaffolded, assignment-specific guidance. Interpretation: student-success programmes should teach students to verify and rework AI output, and should test tutor reliability before recommending a tool. Limits: correlational self-report, synthetic benchmark and qualitative summary; no learning outcomes.
Operating questionDo our AI literacy and tutor-approval processes teach students to rework AI output and check that recommended tools hold correct answers under student pushback?
Supporting evidencePennsylvania State University and Çankaya University (authors)Boston University, Department of Mechanical Engineering (authors)University of Notre Dame (Office of the Provost; Notre Dame Learning Lab for AI in Teaching and Learning)
Evidence ledger
Full record · every source keeps its link and limitations
NYU contract faculty council asks administrators to disable Gemini’s ability to create and edit Google Docs on student accounts, citing revision-history opacity and course AI policies
NYU's Full-Time Continuing Contract Faculty Senators Council passed a resolution earlier in October asking administrators to remove Gemini integrations from Google Docs, Drive and Chrome on NYU accounts. The concern is that students can prompt Gemini outside Docs to create and edit a Google Doc, and if Workspace permissions are enabled the changes appear in revision history as the student's own, undermining a check some professors rely on. The subcommittee chair said the aim is deliberate rather than ambient AI use, not a ban. The council asked for a response on implementation by 1 December. An NYU spokesperson said the AI Leadership Council will review the resolution. NYU has partnered with Gemini since February 2025.
Why it matters, evidence and limitations
- Why it matters
- Shows faculty governance asking for configuration controls on embedded AI in campus productivity suites, a direct constraint on enterprise AI rollouts aimed at students.
- Evidence and measured results
- Reported resolution text summary, faculty and student interviews and a university spokesperson statement; no usage data.
- Limitations and uncertainty
- Resolution is non-binding; technical feasibility of disabling the feature is unresolved; perspectives are from a few faculty and students.
WashU professors diverge on AI: an intro psychology professor replaced writing assignments with in-class exams after colleagues could not tell AI from student writing
WashU's +AI initiative provides enterprise AI tools and a mandatory Canvas module for incoming students. An Introduction to Psychology professor who uses AI for slides and research ran an informal test giving five faculty a mix of AI and student writing; they said they could not identify the origin. He removed writing assignments and now uses three in-class exams and an in-class final, which he called a terrible way to evaluate students, and said he lacks administrative guidance on AI-resistant writing assignments. A creative writing director rejects AI for learning, ignores +AI emails, and starts integrity concerns with a conversation because detection tools are unreliable. Students said instructors should remain the core source of learning.
Why it matters, evidence and limitations
- Why it matters
- A concrete case of AI pressure narrowing assessment to in-class multiple-choice exams, a student-success risk for students who show learning better through writing.
- Evidence and measured results
- Faculty and student interviews; an informal, unpublished faculty detection test.
- Limitations and uncertainty
- Two professors; anecdotal test; reporter is a student in one course (disclosed by the paper).
Notre Dame publishes qualitative findings from its student-faculty AI summit: 88 participants, 1,880+ data points, priorities on motivation, assignment design and course-specific AI guidance
Notre Dame's summit on 14 November 2025 brought together 43 faculty from 36 units and 45 students from 19 majors; more than 1,880 pieces of data were analysed by six trained qualitative coders. Participants wanted to move from preventing inappropriate AI use toward motivating learning, agreeing AI should complement rather than substitute for student thinking. Students named busywork, overwhelming workload, deadlines and grade pressure as reasons for AI use that undermines learning, and participants suggested frequent low-stakes assessments, scaffolded drafts and revision, and in-class work. Students asked for course- and assignment-specific AI guidance rather than one syllabus statement. Faculty asked for time, training and discipline-specific support. Notre Dame will pilot an AI Student Pedagogical Partnership Program this spring.
Why it matters, evidence and limitations
- Why it matters
- Institution-level evidence of what students say drives harmful AI use (workload and grade pressure) and what guidance they want, relevant to student-success policy design.
- Evidence and measured results
- Qualitative analysis of summit outputs by trained coders; summary only, no full methods.
- Limitations and uncertainty
- Self-selected participants at one private university; event nearly a year before publication; summary page without full data; published now because findings are being reported ahead of a November follow-up session.
Preprint: survey of 891 US and Turkish college students finds GenAI experience relates to critical thinking only through dividing labour with the tool and reworking its output
The authors surveyed traditional college students online (917 respondents; 891 analysed: 645 US, 246 Türkiye) using a correlational, cross-sectional design and structural modelling grounded in transactive memory theory. Self-appraised thinking ability and confidence in the human-AI partnership predicted how much practical GenAI experience students built, but neither predicted critical thinking directly. Experience mattered only through two behaviours: dividing and coordinating cognitive labour with the tool, and actively reworking its output rather than accepting it. The structure was largely invariant across countries with two exceptions. The authors acknowledge the design precludes causal inference and that critical thinking enactment was a single dichotomous item.
Why it matters, evidence and limitations
- Why it matters
- Suggests student-success AI programmes should teach how students engage with AI (verification and reworking), not just provide access.
- Evidence and measured results
- Online convenience survey with structural equation modelling and multigroup comparison.
- Limitations and uncertainty
- Preprint, not peer reviewed; preliminary version; self-report; non-US sample requires transferability caution; no learning outcomes.
Preprint: Boston University TrussTutor-Bench tests 13 LLMs on 200 statics problems; models rejected their own correct solutions under student pushback in 1.6% to 50% of cases
The authors built 200 verified two-dimensional truss problems with automatically generated student solutions and evaluated 13 LLMs (ten proprietary from Anthropic, OpenAI and Google; three open-weight) on solving, sycophancy and error detection. Baseline accuracy ranged from 64.5% to 97.0%. Every model detected nearly all injected errors (90.4% to 100%). When a student insisted correct work was wrong, models rejected correct solutions to problems they had themselves solved in 1.6% to 50.0% of cases; confirmation of correct work under pushback ranged from 46.5% to 98.5%, and open-weight models were less reliable on average. The authors advise instructors to tell students that model choice matters and flag that unequal access to reliable models could mean unequal feedback quality. The full run cost about $2,000 in API fees.
Why it matters, evidence and limitations
- Why it matters
- Gives institutions a concrete reliability test (resistance to student pushback) to apply before recommending or licensing AI tutors for STEM courses.
- Evidence and measured results
- Benchmark with verified tasks, structured outputs and per-model accuracy; no classroom deployment.
- Limitations and uncertainty
- Preprint, not peer reviewed; synthetic solutions; single wording; funded in part by an Amazon Research Award and NSF; results do not rank all models.
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.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.