# How Universities Should Actually Respond to Generative AI
## The Real Question Isn’t “To Ban or Not to Ban”
Two major surveys from late 2025 paint a clear picture: generative AI has moved from a novelty to a near-universal presence on university campuses. In one survey of over a thousand full-time UK undergraduates, 95% reported using AI in some capacity for their studies. A parallel voluntary survey at a large American public university drew nearly 95,000 respondents and reached the same figure across 21 different tools. These numbers tell us something important: the institutional debate about whether AI belongs in higher education is effectively over. It is already everywhere students and faculty work.
The more consequential question — and the one that keeps surfacing at major education technology conferences — is where AI genuinely adds value and where it undermines what a degree is supposed to certify. The answer requires universities to make a sharp distinction between two responsibilities they have historically collapsed into one.
The first is teaching students how to work effectively alongside AI tools. The second is independently verifying that students can think, reason, and perform when those tools are absent, unreliable, or inappropriate. Most current policies fail because they never disentangle these two jobs.
## Why AI Tutors Need More Than a Chat Interface
A preprint study published in September drew over 2,300 participants and compared AI-based tutoring against expert human tutoring and no tutoring at all. On surface-level metrics tied to GRE preparation, AI tutoring matched the human tutor group. But the experiment was brief — just one hour — and tested recall immediately afterward. That tells us very little about whether the learning stuck.
A more sobering counterpoint comes from a semester-long trial in an introductory marketing course. There, a retrieval-augmented chatbot carefully integrated into the curriculum produced no measurable difference in student interest, self-confidence, engagement, or grades. The tool was well-built and grounded in the course material. It simply did not move the needle.
What separated the promising experiments from the disappointing ones was design. Dartmouth’s pilot in an introductory statistics course embedded AI into a larger structure of repeated readings, retrieval practice, and AI-graded short answers. The results were striking, showing a large gap between high-use and zero-use groups on final exams. Yet the pilot was optional, observational, and limited to one institution. Self-selection — the students who were already motivated — remains the biggest threat to those findings. Interestingly, the chatbot assistant in that pilot received only 72 total queries, while the structured practice features were used far more extensively.
One more cautionary finding deserves attention. A study from Georgetown and the University of Washington, soon to be presented at a major AI ethics conference, tested what happens when students receive AI-generated summaries of short videos. When the summary was accurate, 83.6% of participants answered a comprehension question correctly the next day. When the summary was misleading, that number dropped to 44.8%. The videos were not course material, but the implication is urgent for any class that allows a summary to substitute for the original reading.
The takeaway is straightforward. Universities should not buy a chatbot. They should buy a learning mechanism with clear goals, inspectable sources, measurable outcomes, and safeguards for the students who fall through the cracks.
## Rebuilding Assessment Around What Students Actually Know
The most practical response emerging from campuses is a redesign of assessment itself. A proposal from Harvard College’s dean outlines a “barbell” model: let AI deepen learning in low-stakes contexts, and make the high-stakes assessments resistant to it through supervised exams, quizzes, and oral defenses for major projects. These ideas are framed as advisory, not yet formal policy, but they point toward a workable middle ground.
The University of Chicago Law School has moved ahead more aggressively. Its strategy for the next academic year includes device-free first-year core classes and in-person exams, combined with a foundation-first approach to legal writing. Students first learn to write without any AI assistance, then progressively use AI for research, revision, and oral-argument preparation. Upper-level research papers add an in-person oral discussion component. The logic is deliberate: some skills must be built without a safety net before the tool is introduced.
A philosophy professor at Williams College has publicly argued that the much-touted evidence for writing as the best path to critical thinking is thinner than advocates claim. He suggests students may develop their thinking more effectively through speaking, answering questions, and defending their positions in real time. Whether one accepts the full philosophical argument or not, the practical point is valuable: an oral defense tests capacities that a polished written essay simply cannot.
AI-permitted work can still demand genuine intellectual effort. A case study from University College London asked students to critically evaluate scientific output generated by ChatGPT. Students reported that the exercise built their independent research skills and gave them a more nuanced understanding of AI’s limitations. The sample was small, but the pattern is instructive: when the model’s output becomes the object of analysis rather than a shortcut to an answer, students engage more deeply, not less.
## The Accountability Gap
The next trap universities must avoid is automating the other side of the desk. Reports from Australia indicate that several institutions now permit limited AI assistance in grading and feedback, with varying safeguards. One university says staff remain fully responsible for marks, while another offers an opt-out for AI-assisted grading. A third explicitly bars AI from assigning any grades. A lecturer at one of these institutions flagged a phenomenon called “verification drift” — where a tired reviewer checks early AI-assisted outputs carefully, then gradually stops scrutinizing them as they become routine.
Faculty norms remain unsettled. A survey of 460 academics who write for publication found that 65% had never used generative AI when preparing their own writing, and roughly a quarter considered it completely unacceptable for academic work. This gap between student practice and faculty norms is already producing visible friction. Incidents involving senior academics using AI on published writing have triggered apologies, and students across multiple campuses have publicly raised concerns about hypocrisy.
The solution is structural, not aspirational. AI can help organize evidence, check a rubric, or surface inconsistencies. But a named academic must own every grade, every piece of feedback, and every misconduct allegation. Students deserve to know when AI has touched their work, to access a route to human review, and to never have an appeal dismissed solely because a detector tool flagged something. The stakes are escalating: reports indicate that students are now bringing legal representation to misconduct proceedings. The cost of getting accountability wrong is no longer hypothetical.
## The Data Question Universities Cannot Ignore
AI policy cannot stop at the classroom door. Recent reporting has revealed that OpenAI used material from the Bodleian Library at Oxford in its training processes. Oxford characterized the use as modest and non-exclusive, restricted to out-of-copyright works, and confirmed that the library retained scan rights and planned open publication of its materials. This was a relatively small footnote in the larger AI debate, but it signals what comes next.
Student work is now squarely in the crosshairs. A major plagiarism-detection company announced plans to use student submissions for AI tool development or improvement. After significant pushback from the UK higher education sector, that change was paused and will not take effect until September 2027, during which time the company is working with a university IT body to redraft the terms. Some institutions have already taken sides: Cambridge is staying on the old contract until July 2027, and Southampton has stated it will not renew after the current academic year. The company behind the plagiarism tool says it does not train a generative model and has no plans to use student essays for that purpose. But a student union officer argued the point more sharply: students cannot meaningfully consent when using the tool is a condition of submitting their work.
The same logic extends to the AI platforms universities themselves are purchasing. One Australian university has provided a campus-wide AI tool to over 80,000 students and staff. A scholar writing in a technology policy publication argues that universities should own the governance layer above any single model and keep their data portable. His litmus test is simple: “Institutions need to be able to leave.”
Access is part of this picture too. One of the largest gifts in higher education history — $3 billion — was announced in late September for a major research university, with a half-billion dollars earmarked for computer science infrastructure including GPUs and AI resources. Most institutions will never have that kind of budget. For them, negotiated campus access agreements with strong privacy defaults and clear exit clauses represent the core equity question of AI adoption in education.
## The Unifying Pattern
What ties together every headline from the past year of AI coverage in higher education? One line: who is accountable for what an AI touches. A student using AI on an assessed paper. A provost using it in published writing. A marker relying on AI to grade student submissions. A plagiarism platform seeking rights to student essays. A research library seeing its collection used for AI training without the kind of explicit consent most people would expect.
In every case, the same question applies. Did a named person make a deliberate decision, and can the people affected see that decision and challenge it? Universities that address this course by course, contract by contract, will keep their degrees meaningful and their communities trusting. Universities that respond with a blanket ban or a single campus-wide license will not solve the underlying problem — they will merely bury it.
## What Institutions Should Do Now
A clear set of priorities emerges from the current landscape:
– Define the specific human capabilities each programme is certifying, and clearly mark which assessments permit AI use and which verify independent mastery.
– Run focused, grounded learning pilots rather than launching campus-wide chatbot platforms. Measure unaided student performance, set clear stopping criteria, and publish what is learned.
– Adopt a human-accountability rule for all grading and misconduct decisions. Require disclosure of AI use, ensure human review is available, preserve the right of appeal, and never treat a detector score as a final verdict.
– Ask faculty and staff the same AI-disclosure questions posed to students, and publish the results transparently.
– Audit the contracts students cannot opt out of — starting with plagiarism checkers and learning management systems — for clauses permitting AI training or data reuse.
– Publish a comprehensive AI procurement register that documents the purpose of each tool, its data flows, training rights, data retention policies, costs, and exit terms.
## Frequently Asked Questions
**Q: Are most universities banning generative AI?**
A: Very few are attempting outright bans, and most of those are struggling to enforce them. The prevailing direction is not prohibition but structured integration, combined with clear rules about where AI can be used and where independent demonstration of skill is required.
**Q: What does “durable learning” mean in the context of AI tutoring?**
A: Durable learning refers to knowledge and skills that persist beyond the immediate test or session. A one-hour tutoring session may produce gains that look good on a post-test taken right afterward, but those gains may fade quickly. True durable learning is demonstrated when students can apply what they have learned days, weeks, or months later, without assistance.
**Q: Why is the summary-substitution problem important for all courses, not just those with assigned readings?**
A: When students receive an AI-generated summary instead of engaging directly with source material, they lose the opportunity to encounter the original framing, tone, nuance, and evidence. The Georgetown and University of Washington study showed that misleading summaries cut comprehension nearly in half. Any course that accepts summaries as a replacement for primary engagement faces the same risk, regardless of discipline.
**Q: What should students do if they believe an AI detector wrongly flagged their work?**
A: Students should know their institution’s appeal process and assert their right to human review. A detector score should never be the sole basis for a misconduct finding. If an institution refuses to provide human review, students may need to escalate through formal complaints channels or seek external advice.
**Q: Can universities really expect students to consent to AI training on their work?**
A: Meaningful consent requires a genuine alternative. When submitting work through a plagiarism-detection platform or a learning management system is a condition of enrollment and assessment, the consent is not truly voluntary. This is why several student unions and policy commentators have called for these contracts to be restructured or for universities to negotiate stricter terms with vendors.
**Q: What is the equity concern around AI access on campus?**
A: Wealthy institutions can invest heavily in AI infrastructure, dedicated tools, and negotiated privacy protections. Smaller and less well-funded institutions may rely on whatever is available, often with weaker privacy safeguards and fewer options for switching providers. The gap risks creating an uneven landscape where students at different institutions receive very different levels of protection and access.
## Conclusion
Generative AI in higher education is not a problem to be solved once and for all. It is a set of ongoing choices about design, accountability, and values. The institutions that navigate this well will be the ones that refuse to confuse a tool ban with a policy, that build assessment around evidence of genuine learning rather than the detection of tool use, and that insist on named human responsibility at every level — from grading to misconduct to data rights. The goal is not to eliminate AI from campus life. The goal is to make sure that when an AI touches a student’s work, a grade, or a research archive, someone is answerable, and the people affected have a voice in the outcome.
Thank you for reading
