- The tool was a GPT-4o "virtual study assistant" inside the LMS that answered from course materials (EdWorkingPaper 26-1598, Annenberg Institute at Brown, dated Sept 30).
- Only about 15% of students offered access used it even once, so the drop is an effect of offering the tool, not of heavy use.
- Estimated losses were larger for first-generation students (about −0.50 SD vs. −0.22 SD for others), though the authors flag limited statistical precision.
- Faculty got only "light training"; a co-author told Inside Higher Ed, "we need to train instructors much more extensively to use these tools."
This Week in 30 Seconds
Four things to know
before your first meeting
The table of contents for a busy week. Read this, then decide whether to keep going.
🎓 Higher Ed · 📊 Research
−0.37
In a University of Maryland randomized trial, offering students a course-specific AI tutor lowered final grades by 0.37 standard deviations, even though only about 15% ever used it.
Source →🎓 Higher Ed
78%
Professors are tracking keystrokes and moving writing back into class; 78% of faculty say AI makes it harder to judge what students know.
Source →🌍 Global · 🛠️ Tools
EU
OpenAI starts watermarking ChatGPT text in the EU, but instructors can't run the detector.
Source →🏫 K-12
0
Common Sense Media rates ChatGPT for Teens an "unacceptable risk," partly because teens can skip its tutoring mode.
Source →🌍 Global · 🎓 Higher Ed · Update
3
Newcastle and Westminster join Southampton in leaving Turnitin (UK). Westminster says the core plagiarism checker is "largely ineffective" against AI.
Source →Main Stories
A tutor that backfired,
and a turn to surveillance
Five developments with enough substance to change what you do this week.
🎓 Higher Ed · 📊 Research · Sept 30 – Oct 7
A randomized trial at Maryland found its AI tutor lowered grades
View generated image — Image source: Nano Banana 2
The University of Maryland tested its own course-integrated AI study assistant before a campus rollout, randomizing 30 instructors teaching 2,379 undergraduates in fall 2025. Among sections of the same course, being offered the tutor reduced final grades by 0.37 standard deviations and LMS participation by 0.90 standard deviations.
Worth noting: With 85% of students never opening the tool, the mechanism is unclear. The authors suggest it may have displaced other course activity or signalled that AI use was permitted, which could change use of ChatGPT and Gemini too. The paper is not yet peer-reviewed; a replication with higher uptake and instructor training would show whether the tool or the rollout caused the drop.
Working paper →
Inside Higher Ed →
What this means for your campus
Before your institution switches on an AI tutor across courses, ask what it will measure and what result would make it switch the tool off.
🎓 Higher Ed · Oct 1–12
Professors are turning to surveillance, and say it is costing them
View generated image — Image source: Nano Banana 2
A Chronicle of Higher Education report (Oct 12) documents faculty tracking keystrokes, hiding prompts in assignments, and moving all writing into class to keep AI out. A National Center for Faculty Development and Diversity survey of 1,659 faculty and administrators finds nearly three-quarters rethinking what they assess, 78% saying it is harder to judge what students know, and only 6.6% saying AI expectations are consistent across their institution. Penn's Marie Gottschalk has dropped take-home papers for the first time in nearly 30 years. Laurel Ridge Community College's Lisa Dunick tracked students' typing in an online summer course and had the lowest success rates of her career; she switched to in-class drafting after reading about AI agents that mimic human typing. Ole Miss's Marc Watkins warns that detection mostly catches students without the resources to evade it, and Tyton Partners finds about a quarter of instructors have moved to in-class proctored exams.
What this means for your campus: Process tracking and in-class writing carry real time and trust costs, so departments should decide together what monitoring they accept rather than leave each instructor to improvise.
The Chronicle → NCFDD report →🌍 Global · 🛠️ Tools · ⚖️ Policy · Oct 5
OpenAI is watermarking ChatGPT text in the EU, but you can't check it
View generated image — Image source: Nano Banana 2
To comply with the EU AI Act, OpenAI will add an invisible watermark, textGrain, to eligible ChatGPT and Codex text generated in the EU over the coming weeks. Detector access is limited to approved researchers and expert organizations. Outside the EU it is opt-in for API customers and off by default. Light edits and copy-paste survive it, but "substantial paraphrasing, or translation" can remove it, and OpenAI warns a missing watermark "does not prove human authorship." Anthropic has announced comparable watermarking for Claude.
What this means for your campus: Watermarks won't settle integrity cases this year: you can't run the detector, and a clean result proves nothing.
OpenAI → Help Center →🌍 Global · 🎓 Higher Ed · Update · Oct 9
Newcastle and Westminster join Southampton in leaving Turnitin
View generated image — Image source: Nano Banana 2
Update on the Turnitin licence story from the Oct 6 issue: two more UK universities are leaving, and one says the core product no longer does the job.
Newcastle will stop after 2026–27, citing the "potential use of colleague and student data for AI purposes." Westminster retires it in July 2027, calling the checker "largely ineffective against the most significant current integrity risk" and its AI detectors "probabilistic rather than conclusive." Aberdeen is reviewing; Cambridge will "not accept any changes to the EULA that impact on data privacy or intellectual property." Southampton paid £87,621 for Sept 2024–Sept 2025.
What this means for your campus: If you renew Turnitin mainly for similarity checking, ask what it still catches when most misconduct involves AI text it can't see.
Times Higher Education →🏫 K-12 · Oct 7
Common Sense Media rates ChatGPT for Teens an "unacceptable risk"
View generated image — Image source: Nano Banana 2
Common Sense Media's Youth AI Safety Institute tested OpenAI's teen experience with more than 4,000 prompts. New parent-linked accounts sent zero alerts after up to an hour of conversation about suicidal ideation, self-harm or disordered eating; alerts appeared only on accounts with weeks of history. A "Show me the answer" option lets teens skip Study mode's tutoring, and adult-registered test accounts never switched to the teen experience, even after the tester said they were 13. The same day, OpenAI announced a College Planner for ChatGPT for Teens to manage application deadlines and financial-aid steps.
Worth noting: OpenAI says "the bulk of their testing may have begun and concluded before activation of parental controls was complete." Common Sense says it confirmed features were live before testing. Retesting on fully activated accounts would settle it.
What this means for your campus: Next year's first-years will have had a one-click way to skip AI tutoring for their answers, so design early assignments that make students show their reasoning, not just the result.
Common Sense Media → CNBC → OpenAI →Research & Practice
What the studies say,
what a colleague tried
Two findings and one classroom move worth borrowing.
From the Research
Research Methods · arXiv preprint · Oct 7
Study participants rarely use AI, but are paid even when they break the rules
What was studied: Researchers linked donated ChatGPT histories to submission records on a high-quality online research platform, covering 712,930 submissions across more than 127,000 studies, plus a survey of 2,500 workers (Sehgal et al.).
What they found: AI assistance appeared in only 1% of submissions, with no rise over three years, and 5% of workers accounted for 64% of it. Assisted submissions were approved at about the same rate as others, even when instructions banned AI. Half involved fixed-choice items that the platform's detector doesn't check.
What you could do with this: If you run online studies, don't assume AI contamination is ruining your data, but don't rely on the platform to enforce a no-AI rule either.
Source →Management Education · ScienceDirect · Oct 8
Business schools describe their AI adoption, but rarely what their degree certifies
What was studied: Weng Marc Lim (Sunway University) reviewed public reports from 84 business schools for evidence that their degrees still certify what graduates can do without AI.
What they found: Schools describe AI access, curriculum integration and faculty development in more detail than any evidence of what the degree certifies. Only one in seven reports a safeguard, such as a defended capstone or work kept free of AI.
What you could do with this: Pick one assessment in your program that could serve as that safeguard, such as a capstone defense or oral exam, and check whether it really is AI-free.
Source →How Others Are Doing It
Indiana University Indianapolis · Africana Studies
Students wrote their own AI policy, then had it stress-tested
Who: Patricia Turley, assistant professor of Africana Studies, IU Indianapolis.
What she did: A three-part assignment. Students wrote an acceptable-use policy with green (permitted), yellow (caution) and red (restricted) tiers, each with three higher-ed examples, benefits, risks and an implementation plan. Two outside readers then "stress-tested" each policy against a rubric, and students wrote a reflection.
What happened: Students voiced concern about their own dependence on technology, and some admitted they weren't following their own rules. One wrote, "I noticed how quickly AI could influence my thinking." These are anecdotal reflections, not measured outcomes.
What you could borrow: Make the policy itself the assignment, and add the outside stress test, which turns a syllabus rule into something students have to defend.
Source →Talk About It
- Maryland's AI tutor lowered grades in a randomized trial. If our institution switched one on, what would we measure, and who would decide to switch it off?
- Faculty are tracking keystrokes and banning take-home writing. How much monitoring of students' writing should we accept before it costs more trust than it protects, and should the department set that limit?
Quick Hits & Radar
Five fast items,
two dates, three actions
Scannable in under a minute, plus what's coming and what to try this week.
- 1Claude now works inside Google Docs, Sheets and Slides A sidebar add-on in public beta on all paid Claude plans edits Workspace files directly (Oct 6).
- 219 of 50 highly selective colleges publish no rule on applicants' AI use An audit by admissions consultancy Oriel finds only Brown, Georgetown and Harvard prohibit AI in application content; just 6 say whether a human or a machine reads applications (Oct 11).
- 3Colleges expand AI security degrees and certificates The roughly three dozen schools in the Pentagon-funded National Cybersecurity Workforce Development Program are building AI-security career pathways (Oct 9).
- 4Knowing the AI rules didn't reduce students' acceptance of misuse In a survey of 304 university students in Kuwait, awareness of AI-plagiarism policy was positively associated with accepting AI-assisted misuse (ECEL proceedings, Oct 7).
- 5OpenAI: 1.2 million teens used Learning Visualizations OpenAI's teen-tier figures also show more than 180,000 teens used Study mode, a sign of how many incoming students already study with ChatGPT (Oct 8).