- RTL found traces of AI in 59% of theses: 14% were more than half AI text, 6% almost entirely AI, and 41% had none detectable.
- Almost half of the largely AI-written theses carried an "AI declaration," and two students who said they used no AI had theses scored 75% and 90% AI. None of the 12 universities kept figures on AI-related thesis fraud.
- Universiteiten van Nederland: "Based on a specific score from a detection tool, it cannot yet be determined whether a student is guilty of fraud," but asked how final assignments show graduates meet a degree's learning objectives.
- Utrecht bans AI-detection software on students' work, Leiden doesn't allow it, and TU Delft calls detector conclusions hard to verify. A week earlier, Amsterdam's Folia found AI in 48 of 50 business bachelor's theses.
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.
🌍 Global · 🎓 Higher Ed · ⚖️ Policy
20%
A TV scan of hundreds of theses from 12 Dutch universities claims about one in five was mostly AI-written. The education minister calls it fraud; the universities say a detector score alone can't prove it.
Source →🎓 Higher Ed
✎
Stanford brings back paper, pencil and weekly reading quizzes as problem-set scores rise, exam scores fall and fewer students come to office hours.
Source →🎓 Higher Ed · 📊 Research · Update
+0.02
A reanalysis of last week's Maryland AI-tutor trial finds no drop in letter grades in the full sample; the headline result doesn't hold up.
Source →🌍 Global · 🏫 K-12 · 📊 Research
~1 yr
PISA: 15-year-olds who use chatbots to summarize, research or draft score about a year of schooling behind non-users, on average. Moderate, purposeful use looks different.
Source →🛠️ Tools · 🎓 Higher Ed
30
OpenAI offers U.S. life-sciences labs up to 30 free seats with its research model; Google opens Deep Review, an AI manuscript checker, to the public.
Source →Main Stories
Detectors on trial,
and a retreat to paper
Five developments with enough substance to change what you do this week.
🌍 Global · 🎓 Higher Ed · ⚖️ Policy · Oct 14–16
A detector scan says one in five Dutch master's theses was mostly AI-written
View generated image — Image source: Nano Banana 2
Dutch broadcaster RTL Nieuws ran hundreds of recent master's theses from 12 universities through the Pangram detector and reported that in about a fifth, chatbots contributed more than the student (Netherlands). Education Minister Rianne Letschert called it fraud; the universities' association replied that a detector score alone cannot establish fraud.
Worth noting: The figures rest on one detector, and RTL did not analyze full theses. Researchers quoted in Dutch campus media say Pangram detects fewer than half of fully AI-written texts, which would make this an undercount; others warn of false positives on individual students. Checking a sample of flagged theses against drafts and oral defenses would settle which way the error runs.
Times Higher Education →
NL Times →
Cursor (TU/e) →
What this means for your campus
If your capstone or thesis is graded only on the written product, plan now for an oral defense or draft review that shows the student can explain the work.
🎓 Higher Ed · Oct 19
Stanford brings back paper as problem-set scores rise and exam scores fall
View generated image — Image source: Nano Banana 2
Stanford's required first-year COLLEGE program is reintroducing handwritten in-class responses, weekly reading quizzes and "screen-free" classes. Vice Provost Jay Hamilton says problem-set scores have improved while final-exam scores dipped, and office-hours attendance is down. Students proposed the quizzes: "Give us low-stakes, understanding checks every week so that we can show that we've been doing the work." One student told him, "Claude is awake at 11:30 at night and you're not. Claude doesn't judge me."
What this means for your campus: A widening gap between homework and exam scores is a cheap early-warning signal; check whether it shows up in your gradebook this term.
Stanford Daily →🎓 Higher Ed · 📊 Research · Update · Oct 15
A second look at Maryland's AI-tutor trial: the grade drop may not be real
View generated image — Image source: Nano Banana 2
Update on the University of Maryland AI tutor trial from the Oct 13 issue: an independent reanalysis challenges its headline findings.
Education professor Brendan Bartanen argues the paper's main claims are "at best, unsupported." The −0.37 SD figure uses Canvas scores in the exact-match sample; in the full sample it is significant only at the 10% level (not the preregistered 5%), and the letter-grade estimate is +0.02 GPA points. The engagement drop mostly reflects control sections assigning more LMS activities, and the first-generation finding is 1 of 4 significant results among 54 untested-for-multiplicity interactions.
Worth noting: The critique is a blog post, not peer-reviewed, and no author response had appeared by Oct 20. Both sides agree only about 15% of students used the tool.
What this means for your campus: If colleagues cite "AI tutors lower grades," the defensible version is narrower: an LMS chatbot that few students used did not help.
Bartanen → Working paper →🌍 Global · 🏫 K-12 · 📊 Research · Oct 13–14
PISA links frequent chatbot use to lower science scores, with exceptions
View generated image — Image source: Nano Banana 2
In PISA 2025, 15-year-olds who use AI chatbots to summarize readings, research new topics or draft writing score about 20 points lower in science than non-users, roughly a year of schooling. Students who use AI to help them learn once or twice a week score about the same as non-users, and frequent users taught at school to evaluate AI output score slightly higher, though advantaged students get more of that teaching. The gap varies: Canada and New Zealand students who regularly draft with AI scored 29 and 36 points lower; in Thailand and Macao (China), 24 and 21 points higher. Across the OECD, 46% of 15-year-olds use chatbots weekly to help them learn.
Worth noting: Weaker students may simply lean on AI more, and PISA can't separate the two. A widely republished piece in The Conversation reads the data as evidence of harm; OECD's Andreas Schleicher says it shows associations only. Longitudinal or experimental data would resolve it.
What this means for your campus: Today's 15-year-olds are your 2029 first-years; whether they were taught to evaluate AI output may matter more than whether they used it.
OECD → The Conversation → Mapping Ignorance →🛠️ Tools · 🎓 Higher Ed · Oct 14–15
OpenAI and Google pitch AI research tools directly at faculty
View generated image — Image source: Nano Banana 2
OpenAI's Rosalind Academic Plan offers U.S. university life-sciences labs up to 30 free ChatGPT Business seats for up to 12 months. A PI or lab lead applies and passes a trusted-access review with identity verification before getting GPT-Rosalind; there is no API access, research data aren't used for training by default, and labs at universities with an OpenAI Enterprise agreement must go through their institution. A day earlier, Google opened Deep Review, which grew out of a tool that gave opt-in pre-submission feedback on more than 15,000 papers at STOC, ICML, NeurIPS and ICLR. In a June preprint it caught 89.7% of known errors on a 29-error benchmark versus 55.2% for a single Gemini response; pilot users also reported it wrongly challenging valid proofs.
What this means for your campus: Before faculty upload unpublished manuscripts or lab data, research offices should check these terms against grant, IRB and data-use agreements.
OpenAI → Help Center → Google Research → R&D World →Research & Practice
What the studies say,
what a colleague tried
Two findings and one classroom move worth borrowing.
From the Research
Assessment · Computers in Human Behavior Reports · covered Oct 13
Students who wrote first, then asked AI what was missing, scored highest
What was studied: University of Eastern Finland researchers analyzed 2023–24 take-home exams from 891 students in an intro qualitative-methods course where AI was allowed with disclosure; 159 used it. They sorted users into handing the task over, using AI for ideas, and writing first then asking for feedback (Hakola, Kukkonen and Valtonen).
What they found: On the 6-point analysis task, hand-over students averaged 3.8, non-users 4.6 and feedback users 5.5; the order held on all three tasks. AI-written answers were fluent but generic and missed the course materials the rubric rewarded. Caveats: about a dozen feedback answers per task, and the study is observational.
What you could do with this: Allow AI on an assignment only after students submit their own draft, and grade explicitly for use of course materials.
Paper → UEF release →Policy · Education Sciences · Oct 2026
University AI policies police integrity; few say how AI should support learning
What was studied: A document analysis of AI policies at 55 universities in six regions, plus 87 institutional, governmental and professional documents collected February 2025–May 2026 (Penev, Slavov and Deliyski).
What they found: Nearly all documents require honesty (97.7%) and academic integrity (95.4%), but only 63.2% require accessibility, 67.8% equity and 9.2% sustainability. 83.9% recognize AI as a supplementary learning tool; just 40.2% recognize AI-supported active learning. It describes policy text, not whether policies work.
What you could do with this: Read your institution's AI policy and note whether it says how AI should support learning, not just what is prohibited.
Paper → Hargis summary →How Others Are Doing It
Harvard SEAS · Applied Physics
A physicist uses AI for the pre-class half of a flipped course
Who: Kelly Miller, Harvard John A. Paulson School of Engineering and Applied Sciences, speaking at the University of Michigan's AI Futures keynote series. She is also chief learning officer of Collage AI, the platform she used.
What she did: Fed her usual pre-class materials and learning objectives into the platform to generate formative assessments and exercises; the platform tutored students through pre-class questions. Instructors choose the assessments and the Bloom's level to target.
What happened: Students valued the feedback, and post-test learning gains were 62% higher than a previous semester. That's a cross-semester comparison reported by someone with a stake in the tool, not a randomized trial; her earlier randomized Harvard physics study found gains roughly double an active-learning class.
What you could borrow: Use AI only for the pre-class step, with formative checks tied to your own learning objectives, and keep class time for practice and discussion.
U-M CAI → Earlier RCT →Talk About It
- The Dutch universities' association asked how final assignments show graduates meet the degree's learning objectives. If a detector flagged a third of our capstones tomorrow, what evidence would we accept that a student did the work?
- A Stanford student said Claude "doesn't judge me" and is awake at 11:30 p.m. What would make our office hours feel lower-stakes than a chatbot, and could students help design that?
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.
- 177% of Russell Group students say they're taught little or nothing about AI Versus 64% at post-1992 universities, in a Bett survey of UK undergraduates; 62% of Russell Group students say what AI teaching they do get is mostly about assignment rules (Oct 15).
- 2Update: Mollick's Co-Existence is out today Answering the Oct 6 radar test, his classroom prescription is "active learning in classrooms, AI tutoring outside classrooms" (Charter interview, Oct 18; New Scientist review, Oct 14).
- 3Universities go on an AI philosopher hiring spree Philosophy of AI is one of the few humanities niches still hiring, including a state-funded AI-and-education cluster at UT San Antonio (Oct 15).
- 4Miami University opens a Center for AI Led by self-described skeptic Dennis Cheatham; Axios also notes the University of Dayton now requires first-years to study AI fundamentals and human dignity (Ohio; Oct 13).
- 530 learning scientists map where AI helps and hurts learning A K–20 expert-consensus report led by Minnesota, UNC and UC Irvine rates 65 learning processes for vulnerability and potential (Oct 8; surfaced this week).
- 6Four states enacted education-specific AI laws in 2026 Idaho, Maryland, Oklahoma and Virginia mostly set guardrails rather than bans, per FutureEd's tracker (Oct 16).