University of Luxembourg

When ChatGPT arrived, universities reached for the alarm—and a moral panic spread fast. Codes of conduct tightened, plagiarism detectors multiplied, surveillance intensified. The threat had a name. The response was swift, punitive—and understandable. Academic fraud has increased since 2022, and the evidence is mounting. The concern is real, and dismissing it wholesale would be dishonest.

But policing alone is not a strategy. It is a postponement.

Because beneath the legitimate fraud problem lies a deeper, more uncomfortable one—a question that generative AI has made impossible to defer any longer: what, exactly, has higher education been valuing?

The Equation that Always Wobbled

For decades, a quiet assumption has underpinned most academic assessment, particularly in humanities and social sciences: produce a coherent, well-informed, well-written text and you have demonstrated that you have learned. This equation was imperfect long before AI arrived. Students could rely on memorisation, borrowed structures and performed erudition to satisfy assessors without genuine understanding. But the equation held well enough, because constructing a convincing simulacrum of thought still required some engagement with thought.

Generative AI dissolved that requirement almost overnight. A system trained on billions of texts can now produce essays that are syntactically fluent, factually grounded and structurally sound—without any process of understanding or meaning-making. To an untrained assessor, the product is indistinguishable. Yet the student’s cognitive process— the struggle, the reformulation, the genuine testing of an idea—is absent or, at best, severely underdeveloped.

"The question was never what can AI do? It was always what have we actually been asking students to do?"

This is not a technology problem. It is an assessment problem that technology has finally forced into plain sight.

What the Mirror Shows 

It would be tempting to conclude that memory, writing and linguistic fluency are therefore worthless—that AI has exposed them as false proxies for intelligence. That conclusion goes too far. These competencies are genuinely valuable, and they must be developed— particularly early in a curriculum, before students have the critical foundations to use AI tools responsibly. A student who has never struggled to construct an argument cannot meaningfully evaluate one that a machine produces for them.

The real problem is narrower but still serious: written output has been used as the primary—and in many contexts the only—evidence of understanding. And AI has revealed that this output can be generated without the cognitive processes it was designed to infer. We were not measuring intelligence. We were measuring one of its surface expressions, and conflating the two.

Cognitive science has long suggested this conflation was a mistake. Genuine intellectual ability is not primarily about retrieval—it lives in adaptation, in generalising from limited examples, in reasoning under uncertainty, in revising a position when evidence demands it. These are the moves that distinguish a learner from a language model. This is not a new diagnosis. The German researcher Axel Krommer argued that schools have long functioned as pedagogical Chinese Rooms: institutions that reward students for executing answer templates and manipulating the right symbols—producing outputs that look like understanding—without genuine comprehension ever being required. Generative AI does not introduce this problem. It reproduces it at industrial scale, and makes it impossible to ignore.

The Redesign That Is Actually Required 

The practical implication is not to ban tools that will define professional life, nor to pretend that anything goes. It is to redesign assessment around what AI cannot replicate: the reasoning behind a position, the capacity to revise under pressure, the ability to defend a choice in conversation.

Process-based approaches—reflective portfolios, annotated drafts, oral defences—make thinking visible in ways that resist easy delegation. Where AI use is permitted, the standard for genuine contribution must rise accordingly. The measure is no longer formal correctness. It is intellectual value added.

This redesign is not cost-free. It demands more instructor time, disciplinary nuance and careful attention to new inequities it may introduce. None of that makes it optional.

The Real Choice 

Higher education now faces a decision it can no longer defer: continue rewarding the simulation of intelligence, or redesign around its genuine exercise. They are not equivalent. One extends a problem AI has merely made visible. The other takes both the technology and the purpose of education seriously.

The question was never what can AI do? It was always what have we actually been asking students to do? It is time to answer it honestly.