- Dartmouth's AAUP chapter is circulating a petition calling for Schnell's removal as provost, citing the college's own disclosure rules for students.
- Beilock wrote that "we need role-appropriate rules and policies for this technology."
- Inside Higher Ed's 2026 survey of 376 chief academic officers found 71% use AI at least weekly and 65% use it to draft communications; only about 1 in 10 institutions has a centralized AI strategy.
- Just 20% of provosts call their approach "primarily generative" (rethinking what it assesses); 8% call it "primarily defensive."
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 · ⚖️ Policy · Update
71%
Dartmouth opens an independent review of its provost's AI use, as a national survey finds 7 in 10 provosts use AI at least weekly.
Source →🌍 Global · 🎓 Higher Ed
125k
Oxford let OpenAI train on 125,000 scanned Bodleian theses. The 2025 partnership announcement never mentioned training.
Source →📊 Research · 🎓 Higher Ed
−13%
Two new working papers disagree on whether AI is already hurting new graduates' first jobs.
Source →🌍 Global · 🎓 Higher Ed · Update
5
Five Australian universities spell out their rules on AI-assisted marking, while Pitt strikes a union clause requiring courses "taught by humans."
Source →Main Stories
A review, a library,
and a jobs dispute
Four developments with enough substance to change what you do this week.
🎓 Higher Ed · ⚖️ Policy · Update · Sept 23–29
Dartmouth orders an independent review of its provost's AI use, and the data says he is not an outlier
View generated image — Image source: Nano Banana 2
Update on the Dartmouth provost story from the Sept 23 issue: the college has now opened a formal review, and a national survey shows how common administrator AI use is.
On Sept. 25, President Sian Beilock announced an independent faculty-and-alumni committee to review how Provost Santiago Schnell used AI "measured against the criteria of the publications to which he submitted his work." The same day, Schnell published a letter saying he should always have disclosed AI use in his public writing.
Worth noting: Much of the case rests on Pangram detector scores. Schnell disputes the "100% AI-written" rating of his Washington Post essay, and detector scores are estimates, not proof. The committee's findings, not the detector, should settle it.
Inside Higher Ed →
Dartmouth →
CAO survey →
What this means for your campus
If your institution requires students to disclose AI use but says nothing about faculty and administrators, expect someone to ask why this term.
🌍 Global · 🎓 Higher Ed · Sept 26
Oxford let OpenAI train its models on Bodleian Library texts
View generated image — Image source: Nano Banana 2
Internal Oxford documents obtained by The Guardian through a freedom-of-information request show Bodleian material digitized by OpenAI was used to "populate the OpenAI training set" (UK). The March 2025 partnership announcement described digitization and access, not training. By June 2025, 125,000 scanned images of 19th- and 20th-century doctoral theses had been shared, along with about 10,000 16th-century ballads. Meeting minutes record governance-committee concerns about reputational risk and environmental commitments. Oxford says the material is out of copyright, "modest in scale," non-exclusive, and will be published openly "in the next few months." Oxford is the only UK member of OpenAI's NextGenAI program, alongside MIT, Caltech and the University of Michigan.
What this means for your campus: Ask your library whether any digitization, plagiarism-checking or licensing deal lets a vendor train on your collections or on student submissions, and whether that was ever disclosed.
The Guardian → Times Higher Education →📊 Research · 🎓 Higher Ed · Sept 10–29
Is AI hurting new graduates? Two working papers reach opposite answers
View generated image — Image source: Nano Banana 2
A U.S. Census Bureau working paper covering about 29% of U.S. bachelor's degrees awarded 2016–2024 finds graduates of the most AI-exposed majors (computer science, information systems, computer engineering) saw initial employment fall 5 percentage points and first-quarter earnings fall 13% after ChatGPT. About half the pay drop comes from lower wages in the same sectors, the rest from moves into lower-wage sectors such as retail; the gap narrows to about 5% after two years. A second paper, from CESifo, using Current Population Survey data, puts summer 2026 unemployment for graduates aged 22–25 at 7.3%, within the 6.3–7.8% range of recent years.
Worth noting: The papers measure different things: first-job employment and earnings by major (administrative data) versus overall unemployment rates (survey data). Neither is peer-reviewed. Applying the Census exposure measure to survey data would help resolve the disagreement.
What this means for your campus: Advisers in computing and accounting should talk about slower, lower-paid first jobs rather than mass unemployment; that is what the evidence currently supports.
Census Bureau → The Chronicle → Ars Technica →🌍 Global · 🎓 Higher Ed · Update · Sept 23–28
AI moves into marking policies and union contracts
View generated image — Image source: Nano Banana 2
Update on Australian AI-assisted grading from the Sept 8 issue: this week's reporting spells out each university's conditions, and a U.S. faculty union put AI into contract language.
Newcastle lets students opt out of AI-assisted marking, Deakin bars AI from "assigning grades," and Western Sydney has joined the permitted group; UNSW, Melbourne and Sydney still refuse (Australia). WSU lecturer Armin Alimardani warns of "verification drift," where markers stop catching AI errors after a run of accurate outputs. At the University of Pittsburgh, the administration's latest counterproposal struck the faculty union's clause that all courses be designed and taught by humans, plus a clause on faculty autonomy over their own AI use. Separately, New South Wales's Education Standards Authority updated its rules so schools cannot rely on AI-detection software to decide cheating cases.
What this means for your campus: Check whether your faculty handbook or contract says anything about AI in grading, course design or evaluation; at Pitt, that silence is being filled at the bargaining table.
The Guardian → Open Campus → The Conversation →Research & Practice
What the studies say,
what colleagues tried
Two findings and two classroom moves worth borrowing.
From the Research
Scholarly Publishing · Organization Science
A journal ran every submission through a detector: 42% more manuscripts, not better ones
What was studied: Organization Science's AI task force, led by Duke's Sharique Hasan, scored 6,957 first submissions and 10,389 referee reports (January 2021–February 2026) with the Pangram classifier and readability indices.
What they found: Submissions rose 42% after ChatGPT, about twice the pandemic-era rise. Only 3.2% of manuscripts scoring 70%+ AI got a revise-and-resubmit, versus 11.9% of low-AI papers. The journal nearly doubled its senior editors to about 60. [Unverified against the primary source: figures as reported by Latin Times.]
What you could do with this: If you edit or review, ask whether your journal's screening capacity has kept pace, and whether your department's incentives reward volume over quality.
Source →Computing Education · arXiv / IEEE ICALT 2026
Most cognitive skills will matter less; critical thinking won't
What was studied: A mixed-methods survey-and-interview study of computing students and professionals on how important cognitive skills were, are, and will be in an AI-rich future (Rani et al.).
What they found: Participants expect most cognitive skills to become less important, with critical thinking the clear exception.
What you could do with this: If your program is revising learning outcomes, ask which ones your students and employers would still rank as essential in five years.
Source →How Others Are Doing It
Christian Kästner · Carnegie Mellon University
Written reflections became 15-minute TA conversations
Who: Christian Kästner, who teaches Machine Learning in Production (17-645, 100–170 students).
What he did: Dropped written reflections. After each assignment, every student has a 15-minute live TA check-in worth 20% of the assignment, pass/fail, with penalty-free retries. Exams rose from 15% to 25% of the grade; an LLM does first-pass homework grading, but no points are deducted without a human.
What happened: TAs spend 50–80% less time grading and more time with students; many students fail their first check-in. He says he has no learning-outcome evidence yet.
What you could borrow: Keep take-home work, but attach a short pass/fail conversation to check understanding live.
Source →School of Nursing · Purdue University
An AI level printed on every assignment
Who: Clinical assistant professors Laura Moffat, Amanda Obourn and Courtney Henchon.
What they did: A five-level scale, from no generative AI to AI-integrated creation, printed on each assignment, with a signed student disclosure. Suspected overuse triggers a non-accusatory review of drafts; a detector score alone cannot establish a violation.
What happened: Faculty say work "felt more student developed" (observational). The open toolkit has been downloaded 167 times in 21 countries.
What you could borrow: The tiers resemble MUSC's framework (Sept 15 issue); the new pieces are the signed disclosure and a fact-finding meeting instead of a detector-based accusation.
Source →Talk About It
- Dartmouth's president says the college needs "role-appropriate rules" for AI. Should the disclosure rule we give students also apply to faculty and administrators, and who would enforce it?
- The Census paper finds graduates of the most AI-exposed majors earn 13% less in their first jobs. Should that change what we tell prospective computing or accounting majors, and who in the department says it?
Quick Hits & Radar
Seven fast items,
two dates, three actions
Scannable in under a minute, plus what's coming and what to try this week.
- 1Google adds study quizzes to Chrome Gemini in Chrome now builds quizzes from an open tab and analyzes non-YouTube audio and video; quizzes start in English on desktop in the U.S. and India.
- 2OpenAI launches an Academy Community Trainer Program Outside organizations can now deliver OpenAI Academy curricula; OpenAI says 4 million people have engaged with Academy content.
- 3INSEAD rolls out ChatGPT Edu institution-wide Students, faculty and staff across France, Singapore, Abu Dhabi and San Francisco get access; no evaluation measures announced.
- 4OpenAI and Greece bring ChatGPT Edu to secondary schools "OpenAI for Greece" is a national-level K-12 deployment other governments may copy.
- 5Anthropic: 68,000 Ghanaian teachers get Claude lesson-planning tools Built with PlayLab and Ghana's Ministry of Education around the national curriculum.
- 6College Board redesigns two AP courses for AI AP CS Principles adds an "AI Create Project"; AP Seminar bars AI while drafting scored essays, shaping the habits your incoming students bring.
- 7A Chicago Booth statistician's 200 AI-assisted papers draw scrutiny Nicholas Polson credits AI for his output; critics quoted by The Washington Post call the pace "unbelievable."