Consider this classroom scenario: seconds after the project is explained, the student already has an answer.

The assignment was to propose a way to make a university walkway safer during heavy rain. Within minutes, an artificial intelligence (AI) tool had produced a polished plan: install sensors, connect them to an alert system, map flood-prone areas and send notifications to students. It was clear, confident and neatly arranged. The student read it aloud.

Then the lecturer asked a simple question: “Who takes responsibility if the alert comes too late?”

The student looked back at the screen. The answer was not there.

That question stayed with me because it points to a different way of discussing AI in higher education. We spend a great deal of time asking what AI can do, and how quickly it can do it. Perhaps we should ask what it cannot carry, and what students must learn to carry themselves.

AI can help us gather information, test ideas and organise a first draft. Used well, it is a scaffold: something that supports a learner while they build understanding. A scaffold helps construction; it becomes a problem when the learner mistakes it for the building.

In the walkway project, the proposal was technically plausible. But the people who used that path were not a set of data points. A student might be rushing to an examination. A cleaner might be pushing a trolley. Someone with limited mobility might have no easy alternative route. A warning could arrive, but would people see it? Would they understand it? Would the route it recommended actually be safe?

These are questions of context and consequence. They ask us to look beyond whether a system works in theory and consider what happens when it meets real lives. AI can suggest possibilities, but human judgement has to weigh them. Empathy asks us to notice who may be overlooked. Ethical reasoning asks whether a decision is fair, even when the easiest solution is cheaper or faster.

And then comes accountability.

If the plan fails, the software does not stand before the people affected and explain what happened. It does not accept responsibility for approving the budget, choosing the route or deciding that a warning was good enough. People do. Students should learn that responsibility cannot be handed to a tool just because the tool helped produce the answer.

This is why ownership matters. A student who submits AI-generated work without understanding it has not gained a shortcut to competence. They have transferred the appearance of competence onto a page. Proactive ownership means asking where the answer came from, checking what it leaves out, making a decision and being prepared to defend that decision.

Higher education should therefore give students harder problems. Not because they require more pages, longer formulas or another late night, but because they are uncertain, connected to people and resistant to one clean answer. Students should work with incomplete information, interview people who use a space, test prototypes, learn when an idea fails and revise it. They should explain not only what they made, but what they chose, whom it serves and what could go wrong.

A difficult project should not be a trap. It should be a space where students can fail safely, learn openly and try again. The point is to make the student more capable because AI is available.

We must also be careful about what we mean when we talk about original thinking. Not every valuable idea arrives by following the usual route. Some students notice patterns others miss; some challenge assumptions because the standard explanation does not fit how they experience the world. Neurodivergent students, among others, may bring perspectives that a conventional classroom can overlook. They should not be asked to think alike in order to prove they can think critically. A learning environment should make room for different ways of seeing the problem.

AI in higher education places a new pressure on all of us: return to first principles. When a tool can produce a fluent answer in seconds, students need to know what makes an answer sound, what evidence supports it and which assumptions hold it together. They need to understand the fundamentals well enough to spot a confident mistake. Without that grounding, a polished response can become a hiding place.

The student with the walkway proposal did not need a lecture about rejecting AI. The tool had helped the group move quickly. What the student needed was the next question, then the next decision. Who uses this path? What happens when the system is wrong? Who checks it? Who is accountable? By the end of the discussion, the original proposal had changed. The group had not simply added more technology. It had begun to understand the problem.

That is the kind of learning we should protect. AI can scaffold the work, but it cannot take ownership of the consequences. It can offer a route through the problem, but it cannot decide what we owe to the people waiting at the other end.

The future of higher education may depend less on teaching students to compete with machines and more on teaching them to think, choose and answer for what they do. The difficult question is no longer whether AI can produce an answer. It is whether the student can stand behind it.