- Cornell users must complete an introductory training on AI and responsible use before getting the advanced tools. It is a two-year pilot, and faculty and staff can buy extra capacity at a discount.
- Cornell picked several vendors "because no single AI model is best for every discipline or task," and its contracts bar training on what users enter.
- Princeton faculty may require students to use the campus ChatGPT on assignments that allow AI, "and that students include all logs with their submission." Faculty Pro accounts carry a $50 monthly departmental charge. Historian Kevin Kruse compared the giveaway to "unlimited heroin or free handguns."
- Campus deals don't lock in capacity: Stanford told ChatGPT Edu Pro users that OpenAI's DevDay usage cuts apply to them, and users who had used half their weekly allowance saw the rest drop to zero.
This Week in 30 Seconds
Five 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 · 🛠️ Tools
3 vendors
Cornell will give faculty, staff and Ph.D. students free ChatGPT, Claude and Gemini from November, with all students by early 2027. Princeton gives every student ChatGPT Plus from January and lets faculty require the chat logs.
Source →🎓 Higher Ed · ⚖️ Policy
224
An AI tutoring contractor emailed SUNY Geneseo students without the provost's approval; she had to restate that each instructor sets the AI rules.
Source →📊 Research · 🌍 Global
~0
An audit finds the meta-analyses behind "AI boosts learning" badly flawed; a re-analysis of 44 well-designed experiments averages roughly zero.
Source →🌍 Global · ⚖️ Policy
March
Australia's University of Tasmania will have its assessments audited so its degrees can be certified as "personally earned."
Source →🎓 Higher Ed
✓
A Fields medalist at UC Berkeley now grades homework on completion only, because AI makes correct answers meaningless as evidence.
Source →Main Stories
AI for everyone,
evidence for no one
Five developments with enough substance to change what you do this week.
🎓 Higher Ed · 🛠️ Tools · Oct 26–28
Cornell and Princeton put campus AI accounts in every student's hands
View generated image — Image source: Nano Banana 2
Cornell will give faculty, staff and Ph.D. students free base-level access to ChatGPT Edu, Claude Enterprise and Gemini starting in November, with master's students and undergraduates following in early 2027. On October 28 Princeton announced a one-year pilot giving every student ChatGPT Plus from January.
What this means for your campus
Once every student has the same licensed tool, your syllabus policy becomes the main variable. Decide before spring whether you will require the campus account, and whether you will collect logs.
🎓 Higher Ed · ⚖️ Policy · Oct 22
An AI tutoring vendor emailed SUNY Geneseo students, and the provost had to restate whose rules apply
View generated image — Image source: Nano Banana 2
On October 19 the CEO of CircleIn, a company under contract with the college, emailed 224 students promoting its AI tutor ("Ever stared at an assignment and thought, 'I have no idea where to start'?"). Faculty were angered; the Chronicle's headline summarized it as "Don't Worry About Professors' Rules." Provost Mary Toale said she hadn't approved or known about it, required clearance for future messages, and had the company tell all 224 students that professors' AI policies apply in every course. The CEO called it "a mistake that was resolved." Rutgers' Britt Paris says colleges push through contracts for "unproven technologies" without proper vetting.
What this means for your campus: Vendors with student email lists can send students messages about AI that contradict yours. Ask who approves vendor messages to students.
Chronicle → Daily Briefing →📊 Research · 🌍 Global · Oct 26–27
An audit says the meta-analyses showing AI boosts learning don't hold up
View generated image — Image source: Nano Banana 2
Ivy Tech professor Patrick O'Neill audited 14 of 39 meta-analyses reporting big AI benefits and found "impossible statistics," "hallucinated references" and comparisons that don't test AI (e.g., AI-supported study sessions vs. no study at all). The 14 have 3,000+ citations; one has been retracted, and two-thirds of journals' calls for papers invited only research on AI's benefits. His preliminary re-analysis of 44 well-designed experiments averages roughly zero. "All of these meta-analyses are flawed," said MIT's Justin Reich; Stanford's Susanna Loeb found only 20 of almost 900 studies met her team's bar.
Worth noting: A new Frontiers meta-analysis of 58 studies reports g = 0.864, but its prediction interval runs from −0.21 to 1.94 and small-study correction shrinks it to 0.26–0.45. O'Neill's re-analysis isn't yet peer-reviewed; publishing its methods would let others check the near-zero average.
What this means for your campus: When someone cites "meta-analyses show AI improves learning," ask what the comparison group actually got, and pilot with a small group first.
FutureEd / Hechinger → Audit → Frontiers →🌍 Global · ⚖️ Policy · Australia · Oct 27
An Australian university will have its degrees audited as "personally earned"
View generated image — Image source: Nano Banana 2
The University of Tasmania is applying an "assured assessment standard" and aims for every course to meet it by the end of March. Proudly Human, a start-up founded by former chief scientist Alan Finkel, will audit course designs, sample assessments and interview a few students and staff, like a financial audit. Students can still do AI-permitted work, but must pass the audited, AI-free assessments to qualify. The standard builds on Sydney's "two-lane" model and is open to other institutions. Finkel's pitch: employers now assume "everybody's cheating."
What this means for your campus: Two-lane assessment is starting to be audited from outside. List which assessments in your program would count as the secured lane.
Times Higher Education → The standard →🎓 Higher Ed · Oct 23–27
A Fields medalist at Berkeley stops grading homework for correctness
View generated image — Image source: Nano Banana 2
Mathematician Richard Borcherds now credits homework for completion, with final grades resting mainly on in-person exams. "AI can now solve many undergraduate mathematics problems, often better than the students themselves," he told Newsweek. Grading for correctness, he says, "penalizes students for honesty." He calls his approach "not a very good system" but knows no better. Berkeley asks faculty to state AI expectations in syllabi but leaves the rules to instructors.
What this means for your campus: If homework is mainly practice, grading it on completion and moving points to in-person checks removes the reward for outsourcing it, though motivated students may still skip the practice.
Daily Californian → Newsweek →Research & Practice
What the studies say,
what colleagues tried
Two findings and two classroom moves worth borrowing.
From the Research
Cognition · Information and Software Technology · Oct 22
STEM students who trust AI most report the least cognitive engagement
What was studied: Oregon State researchers surveyed 299 STEM students at five North American universities about AI trust, routine use and cognitive engagement (Choudhuri, Sarma, Burnett and Sanchez).
What they found: Students who trusted and routinely used generative AI reported significantly lower cognitive engagement. The authors describe a "self-reinforcing cognitive debt cycle," with the most AI-savvy students deepest in it. Caveat: a one-time, self-reported survey shows association, not cause.
What you could do with this: Pair AI-assisted problem solving with a supervised step where students critique, debug and verify the AI's work without AI.
Paper → OSU release →Learning analytics · arXiv preprint · Oct 24
A semester of AI-tutor chat logs barely predicted students' grades
What was studied: Whether 16,851 student–tutor exchanges from two offerings of an undergraduate AI course (175 students with matched outcomes) predict assignment scores better than guessing the class median (Dong et al.).
What they found: The best model's average error was 3.02 grade points vs. 3.14 for the median guess; the confidence interval for the gain included zero. Errors were largest for the lowest-scoring quarter, the students an early-warning tool most needs to catch. Scores bunched near the top (91% above 90).
What you could do with this: Don't count on chatbot logs to flag struggling students; keep a short, human-reviewed check-in early in the term.
Paper →How Others Are Doing It
University of Virginia · First-Year Writing
A writing course bans AI, and explains why in terms students hadn't considered
Who: Ethan King, who teaches first-year writing at UVA, as described by Derek Bruff after his workshop on AI syllabus statements (Oct 23).
What he did: Bans AI in a seminar on attention and distraction, with a rationale: writing "only works if the seeing, thinking, and feeling is yours to begin with. AI has nothing at stake in what it writes. You do." He also teaches a separate course on writing in the age of generative AI.
What happened: King says this view of writing is new to many students, who focus on the finished product, which is exactly what makes AI tempting. Bruff frames it as unlearning. Instructor observations, not measured outcomes.
What you could borrow: Whatever your AI rule, write down the belief about your discipline it rests on, and name the habit students need to unlearn.
Derek Bruff → Full statement →Cornell · Architectural History
Students argue with chatbot versions of the king and Parliament
Who: Pamela Karimi, associate professor in Cornell's College of Architecture, Art and Planning and one of 33 AI Experimenters Faculty Fellows (Oct 22).
What she did: Students take the role of John Evelyn, whose 1661 "Fumifugium" proposed clearing London's smoky air, and debate source-grounded chatbots playing King Charles II and Parliament.
What happened: Karimi says students better understood the political and environmental obstacles Evelyn faced, and the bot's historical errors became discussion material. No outcome data reported.
What you could borrow: Cast the chatbot as an opponent the student must persuade, not an answer source, and treat its mistakes as material for close reading.
Cornell Chronicle →Talk About It
- Cornell and Princeton are giving every student the same licensed AI tool. If our campus did that, should our course policies assume students use it, and what would we stop grading as a result?
- Borcherds says grading homework for correctness "penalizes students for honesty." Which graded assignments in our program mostly reward whoever used AI, and should they still carry points?
Quick Hits & Radar
Six fast items,
two things to watch, three actions
Scannable in under a minute, plus what's coming and what to try this week.
- 1Springer Nature and Google DeepMind are building an AI peer-review assistant A Gemini-based prototype that helps reviewers analyze text, charts and figures is in testing inside Springer's private environment (Oct 26).
- 2Google Classroom teachers can now read students' Gemini Notebook chats AI summaries of recurring questions plus full transcripts on assigned notebooks; students are notified; Gemini LTI support for LMS users is due in coming weeks (Oct 27).
- 3AI majors nearly doubled in under a year 107 U.S. and Canadian universities now offer one, up from 57 last December; MIT's AI and Decision Making is its second-most popular major (WBUR, Oct 23).
- 4College presidents call AI "confusing" and "chaotic" At Inside Higher Ed's summit, some admitted having no AI policies, and several avoid committing to a single platform (Oct 22).
- 5House of Lords Library asks whether AI is weakening the value of a degree The UK briefing notes the London School of Innovation now offers master's programmes taught mostly by AI tutors (Oct 27).
- 6Concordia College plans an AI major and campus-wide faculty training A $3 million, five-year U.S. Education Department grant will train faculty to bring AI into all majors (Minnesota; Oct 27).