Universities cannot grade their way out of AI  October 4th 2026

TL;DR

• Almost every student now uses AI. In a UK survey of 1,054 undergraduates, 94% said they used generative AI for assessed work, yet only 48% felt staff were helping them build AI skills. EDUCAUSE's 882 respondents just made "determining where AI adds real value" the sector's top IT issue, ahead of cybersecurity for the first time.
• AI tutoring works only with a learning design. A GRE preprint found AI tutoring matched expert human tutors in a one-hour session, while a semester-long course chatbot changed nothing measurable for roughly 500 undergraduates.
• Universities are splitting assessment into two lanes: AI-enabled work, plus supervised exams and oral defences that verify what students can do alone.
• Faculty are now in the spotlight too. Dartmouth is investigating its provost after the student paper found writing that appeared AI-generated, and students are hiring lawyers to fight AI-misconduct accusations.
• Student work is becoming training data. Cambridge refused to sign Turnitin's new licence, which would have allowed student submissions to be used to develop AI tools; the change is now postponed to September 2027.

1. The argument is over. The design problem is not.

In a December 2025 survey of 1,054 full-time UK undergraduates, 95% said they used generative AI in at least one way. The share who said they directly included AI-generated text in assessed work reached 12%, up from 8% a year earlier. The figures are self-reported, UK-specific and from a survey sponsored by an education-technology company. But a much larger voluntary California State University survey points the same way: more than 94,000 students, faculty and staff responded, and 95% had used at least one of 21 listed AI tools.

The people who run campus technology have reached the same conclusion. EDUCAUSE released its 2027 Top 10 IT Issues at its annual conference in Denver, which closed on October 2. The top issue was not "deploy more AI." It was deciding where AI is worth it.

So the institutional choice is not "AI or no AI." It is how to separate two jobs universities have let blur together:

1. Teach AI-enabled work. Students need practice using, questioning and supervising the tools they will meet after graduation.
2. Verify independent capability. A degree still has to mean its holder can reason, explain, decide and perform when the tool is unavailable, wrong or inappropriate.

2. An AI tutor needs a learning design, not just a chat box

StudentBench, a September preprint surfaced in AI Weekly's live feed, assigned 2,383 participants to AI tutoring, expert human tutoring or no tutoring. The authors report that AI tutoring produced statistically equivalent GRE learning gains to expert human tutors. But the task was GRE preparation, the session lasted one hour and the post-test was immediate. It does not establish durable learning.

The peer-reviewed counterweight is a semester-long trial in introductory marketing, where a course-grounded, retrieval-augmented chatbot produced no statistically significant change in interest, self-efficacy, engagement or achievement.

Dartmouth's Phosphor pilot hints at what separates the two. It built readings, repeated retrieval practice and AI-graded short answers into an introductory statistics course. The workshop paper reports a modeled full-versus-zero-dosage gap of 0.71 to 1.30 standard deviations on the final exam, but the deployment was optional and observational at one selective institution, so self-selection is the central threat and the gap is not a causal effect. Its chat assistant got only 72 queries; the structured practice features were used far more.

There is also a reason to be careful about the most common student use of all: summarising. A Georgetown and University of Washington study, to be presented at the AAAI/ACM AI, Ethics and Society conference this month, showed 331 participants a short animated video and, a day or two later, gave them an AI summary of it. On a key question about what they had seen, 83.6% of those given an accurate summary answered correctly, against 44.8% of those given a misleading one. The clips were not course material, but the warning applies to any course that lets a summary stand in for the reading.

The lesson is not "buy a tutor." It is "buy a learning mechanism." Ask what students will do differently, whether faculty can inspect the system's sources, whether the pilot measures unaided performance rather than clicks, and who is helped or left behind.

3. Build assessment around evidence of learning, not detection

Harvard College's dean has proposed a “barbell” approach: encourage AI where it deepens learning, and make mastery assessments resistant to it, with more supervised quizzes and exams and oral defences for big projects. He is explicit that these are advisory ideas, not yet policy.

The University of Chicago Law School has gone further. Its 2026–27 strategy combines generally device-free first-year core classes and in-person exams with foundation-first legal writing: students learn to write without AI, then use it for research, revision and oral-argument preparation. Upper-level research papers add an in-person oral discussion.

That shift has a philosophical defender. In a New York Times guest essay, Joe Cruz, who chairs philosophy and cognitive science at Williams College, argues that the evidence for writing as the best way to develop thinking is thin, and that students may think best by speaking, answering questions and defending their ideas. You do not have to accept his whole case to take the practical point: an oral defence tests something a polished essay no longer can.

AI-permitted work can still demand real thought. A UCL case study asked students to critically evaluate ChatGPT-generated scientific output; students said it built independent research skills and an understanding of AI's limits. It is one small cohort, but the pattern is useful: make the model's output the object of judgment, not a substitute for it.

4. Keep a named human accountable for marks and misconduct

The next temptation is to automate the other side of the desk. A story shared in AI Weekly's expert pool found that several Australian universities permit limited AI assistance in assessment or feedback, with caveats: Western Sydney says staff remain responsible for marks, Newcastle offers an opt-out, and Deakin bars AI from assigning grades. A Western Sydney lecturer warns of "verification drift," where an overworked reviewer checks early outputs carefully and slowly becomes a rubber stamp.

Faculty norms are not settled either. In a Chronicle of Higher Education survey of 460 academics who write for publication, 65% said they had never used generative AI when preparing writing for publication, and a quarter called it "completely unacceptable" for academics. The gap is already producing incidents. Dartmouth's provost, Santiago Schnell, has apologised, and the Times reports students on several campuses calling such cases hypocrisy.

The rule should be simple. AI may help organise evidence or check a rubric, but a named academic owns the grade, the feedback and any misconduct allegation. Students should know when AI touched their work, have a route to human review, and never lose an appeal to a detector score. The cost of getting this wrong is rising: the Financial Times reports students are now bringing lawyers to misconduct cases.

5. Universities are becoming data suppliers, too

AI policy cannot stop at the classroom. Guardian reporting says OpenAI trained on Bodleian Library material. Oxford said the use was modest, non-exclusive and limited to out-of-copyright works, and that the library kept scan rights and planned open publication.

Student work is next in line. Turnitin planned to let student submissions be used for "improving or developing AI tools." After objections from the UK sector, the change was paused until September 2027, to be redrafted with the university IT body UCISA. Cambridge stays on the old terms until July 2027; Southampton says it will not renew after 2026–27. Turnitin says it does not train a generative AI model and has no plans to use student essays that way. As a York Students' Union officer argues in Wonkhe, students cannot meaningfully consent when using the tool is a condition of handing in their work.

The same logic applies to the AI platforms universities are buying. UNSW Sydney has given ChatGPT Edu to more than 80,000 students and staff. In Tech Policy Press, Tom Smith of the Royal Air Force College argues universities should own the governance layer above any single model and keep their data portable. His test: "Institutions need to be able to leave."

Access is part of this too. On September 30, Carnegie Mellon announced a $3 billion commitment from Ken Griffin, which it calls the largest in higher-education history, including $500 million for its School of Computer Science that will pay for GPUs and AI infrastructure. Most universities will never have that budget. For them, negotiated campus access, with privacy defaults and an exit clause, is the equity question.

The pattern behind the headlines

Every story this week is about the same missing line: who is accountable for what an AI touches. Students using AI on assessed work, a provost using it on published writing, an Australian marker using it on grades, Turnitin seeking rights to student essays, and Oxford sharing library scans all raise one question. Did a named person decide, and can the people affected see and challenge that decision? Universities that answer it course by course and contract by contract will keep their degrees credible. Those that answer it with a ban or a campus-wide licence will not.

What to do this week

• Name the human capabilities each programme certifies, and mark which assessments allow AI and which verify mastery without it.
• Run one grounded learning pilot instead of a campus-wide chatbot launch. Measure unaided performance and set a stopping rule.
• Adopt a human-accountability rule for grading and misconduct: disclosure, human review, appeal and a named decision-maker. A detector score is never enough on its own.
• Ask staff the same AI-disclosure question you ask students, and publish the answer.
• Audit the contracts students cannot opt out of, starting with plagiarism checkers and the learning management system, for AI-training and data-reuse clauses.
• Publish an AI procurement register with purpose, data flows, training rights, retention, cost and exit terms.

Wait, What?

An Australian tutoring company told parents to save their money and use AI instead. Dymocks Tutoring and Talent 100 closed its five Sydney centres at the end of September, telling customers that tools like Gemini and ChatGPT had made its service obsolete. It is rare for a business to recommend its replacement on the way out. Universities, which sell much more than tutoring, should read it as a question: what do we offer that a chatbot cannot?

Worth Watching

• MIT's August report on AI and education: a faculty committee finds AI is pulling students away from study groups and office hours, and recommends redesigning courses from their learning goals, with oral exams and portfolios, rather than trying to AI-proof them. Shared by 10 experts in our network.
• Brown's four-year study of an AI student-support system: text-message nudges improved completion of time-sensitive admin tasks but had no detectable effect on grades or persistence. A useful guide to where campus agents belong.

This week’s poll

Where should your university draw its firmest AI line?

— Alexis