The Cornell Paper Is Not a Cheating Story

A new paper landed in Science on 21 May 2026 — René Kizilcec and colleagues at Cornell, Generative AI Use and Misuse Call for Assessment Reform in Higher Education. The number being repeated in the press is the easy one: about a third of US undergraduates regularly use generative AI on assignments, 9 percent admit to using it to cheat. Reporters picked the cheating number because cheating is a story people already know how to read.

That is not what the paper is about.

The paper is about whether a degree still maps to a demonstrated capability. Strip away the moralising about ChatGPT, and the underlying claim is that the artefacts students hand in — essays, problem sets, code commits — have decoupled from the cognitive work the artefacts were always a proxy for. The diploma is a contract between an institution and an employer, mediated by grades that signal something about the holder. If the signal has been polluted, the contract is in trouble. Kizilcec said it plainly: "the credibility of university credentials" is the problem.

Read the methodology before reading the conclusion. The Cornell sample is roughly 95,000 students at 20 US public research universities, surveyed on their own AI behaviour. That is self-reported data, of the same family the Microsoft Work Trend Index and the McKinsey State of AI sit in — the kind I spent a full piece picking apart less than a month ago. Students who admit on a survey to using AI on an assignment are a floor, not a ceiling. The 9 percent cheating figure is what young people will tell a researcher. The honest number is higher, and the survey instrument is incapable of finding it. The honest measurement problem is everywhere; it does not stop at the campus gate.

what the regulators were already trying to do

While the Science paper was in press, the European Commission's apparatus had already started moving on the same problem from a different direction. Article 4 of the EU AI Act has been in force since 2 February 2025, requiring every provider and deployer — including universities deploying GenAI in teaching — to ensure their staff and "other persons dealing with the operation" of AI systems have a sufficient level of literacy. The fines for non-compliance reach €7.5M or 1 percent of global annual turnover. That clause is doing real work in Member State guidance memos, and it lands on university procurement offices whether or not they have noticed.

Sitting underneath the regulation, the operational push: the Commission has carved out a €7M call within the Digital Europe Programme for a sectoral Digital Skills Academy in GenAI, with three more academies in quantum, semiconductors and virtual worlds at €12.5M combined. Alongside it, the PANORAIMA consortium — sixteen partners, eight universities, four research centres, four SMEs — runs its first pilot specialisation tracks this September, with full module availability projected for September 2027. PANORAIMA is the follow-on to HCAIM, which closed in 2024 after delivering a 60-credit human-centred AI master's at BME, UPC, NCI, Sapienza and Utrecht. I have written about both before; my interest here is the timing.

Roughly twelve weeks separate the September PANORAIMA pilot from the new academic year that Science has just told American deans is structurally compromised. The European response is on schedule and arrives about six months after the diagnosis. That is unusual for European institutional speed. It is also the reverse of the usual transatlantic pattern, where US universities iterate quickly while Brussels stays in committee. Worth watching whether the September cohort produces useful signal or merely useful press.

the vendor solution that wants to be the answer

The other timeline that crosses Cornell's is Anthropic's. Claude for Education now has campus-wide access agreements with Northeastern, LSE, Champlain, Syracuse, Dartmouth, the University of Virginia, the University of Pittsburgh, Northumbria, and the University of San Francisco. The product's centrepiece is "Learning mode", which is meant to nudge a student toward Socratic engagement rather than answers. The same month the Cornell paper came out, Anthropic also opened a Claude Corps fellowship and student-builder programme on the campus side. Google opened an AI fellowship for university faculty earlier in the year. OpenAI launched ChatGPT Edu into the same hole.

The vendor framing is: campus-wide licence plus a learning-mode toggle plus a faculty grant programme equals institutional AI literacy. Treat that skeptically. A toggle that asks Socratic questions does not change the underlying assignment design, the rubric, the gradebook, the registrar, or the credential. Toggles are a layer on top of an unreformed substrate. The Brookings report A New Direction for Students in an AI World, the one that called generative AI "the fast food of education", reviewed more than 400 studies and concluded that the risks of overreliance currently outweigh the benefits, especially in foundational-knowledge acquisition. Frictionless is the problem, not the solution. A learning-mode toggle keeps the friction theatrical.

I do not want to be unfair to Anthropic — Learning mode is a more thoughtful product than most of what shipped into education this year. But pricing your way into a campus does not constitute pedagogical reform, and a chatbot vendor cannot replace the institutional decision about what a graded assignment is meant to certify. That decision belongs to faculty, with the registrar in the room.

what the reform actually costs

Assessment reform is the unglamorous category, and it is what the Cornell authors are asking for. In practice it means in-person oral defences, supervised problem sessions, project iteration with version-controlled diffs that show the human's hand, applied coursework graded on process not artefact, and a deliberate move away from the take-home essay as the dominant signal. Each of those modalities costs faculty time. Each requires staffing levels that twenty years of higher-education austerity have systematically removed.

The OECD's Digital Education Outlook 2026 names the same trade in less inflammatory language: generative AI can support learning when guided by clear pedagogical principles; without that scaffolding it inflates output without improving cognition. That is the line vendors quietly elide and that public-policy people will have to underwrite. Pedagogical scaffolding is not free. Somebody — the ministry, the endowment, the student through tuition — is going to pay for it, and the bill is showing up at the same moment most higher-education systems are in fiscal compression.

There is one more cost nobody wants to talk about. A fraction of the current undergraduate cohort already has a fluent, instrumented relationship with AI tooling that their professors do not have. A fraction does not. Assessment reform that ignores this asymmetry will entrench it. Reform that addresses it has to teach AI use as a graded competency, not as something that contaminates assignments. That is the design problem PANORAIMA is now actually trying to solve, and it is the design problem Article 4 quietly forces every European employer to face by way of literacy obligations on its own staff. The campus and the workplace are converging on the same problem from opposite ends.

the stake

Two things are being sold to universities right now as if they were the answer, and both are wrong.

The first is the campus-wide LLM licence, filed as institutional AI strategy. It is a procurement decision, not a strategy. The second is the AI-detector arms race, which has been shown to misfire against non-native English speakers and is generating its own pipeline of false-accusation lawsuits. The combined effect is that institutions pay twice — once to enable AI use, again to surveil it — without changing the assignment design that pollution problem began in.

What the diagnosis actually demands is three things, none of which are quick. A modest, expensive, in-house assessment-redesign programme, owned by faculty rather than the EdTech office, built on the in-person modalities described above. A clear graded curriculum in AI use, modelled on what the HCAIM consortium proved could be built and what PANORAIMA is now scaling — not as a separate certificate but as a layer running through every degree. And a single sentence in the next graduation programme that says plainly what the diploma attests to in 2026. Honesty about the contract is the precondition for repairing it.

The Cornell paper will be read as a cheating story all summer. It is not a cheating story. It is the first peer-reviewed admission, in a journal of record, that the most important thing universities sell has lost some of its meaning. The institutions that figure out how to put the meaning back will be the ones worth attending in 2030. The rest will keep selling a credential the employer market has already begun to discount.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan, an energy-AI venture, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.