AI, Essays and University Assessment: How Do We Know What a Student Actually Knows?

Katherine Rundell’s recent Guardian essay, “I hate what AI is doing to the minds and happiness of the young”, made me think about a distinction that is getting lost in the debate around AI and education.

Rundell, a children’s author and Oxford fellow, argues that generative AI risks weakening the intellectual habits education is supposed to develop: reading difficult material, tolerating frustration, forming arguments and learning to think independently. She describes students using ChatGPT for homework and university essays, and makes a particularly sharp observation about what has changed. Students have always been able to write an essay about a book they had not properly read. Now they can submit an essay they have barely read either.

That is a real problem.

I disagree, though, with some of the conclusions we are drawing from it.

One of the worries running through Rundell’s article is that easy access to AI encourages intellectual surrender: students give up earlier, do less of the difficult work and may mistake fluent output for their own understanding. She cites research on reduced persistence during AI-assisted problem-solving and argues strongly for the value of struggle in learning.

I find the idea that AI will make people fundamentally less interested in knowing something less convincing.

People who have a genuine drive to understand tend to pursue knowledge because they want the knowledge itself. Having a machine capable of producing a plausible answer does not remove that desire. If anything, a curious person now has an extraordinarily fast opponent, tutor, research assistant and source of questions available to them.

Other students were never particularly interested in mastering the subject. Education has always contained people doing the bare minimum to pass.

AI changes something else. It makes it much easier for those students to look as though they know.

AI has damaged an old proxy for knowledge

For a long time, teachers could use completed work as a reasonably useful proxy for the work that had happened inside a student’s head.

It was never perfect. Students copied homework, borrowed ideas, skimmed texts, received too much parental help and occasionally produced essays about books they had barely opened.

Generative AI changes the scale of the problem.

A student can now produce an articulate, structured piece of writing on a subject they understand very poorly. With enough editing, it may be extremely difficult for a teacher to establish from the text alone how much of the intellectual work belongs to the student.

Rundell makes exactly this point. She says academics who believe they can identify AI-written work may be overestimating their ability to do so; poorly used AI is relatively easy to spot, whereas carefully edited AI-assisted writing can be much harder to distinguish.

I agree with her there.

Where I steer away from the more defensive response to AI is what we do next.

Trying to make every homework assignment AI-proof seems close to impossible. Trying to determine retrospectively exactly how much AI entered a piece of unsupervised work may become increasingly futile.

Perhaps we should stop asking those pieces of work to prove so much.

Essays should remain part of learning

Rundell defends the essay for a reason I agree with. Writing is itself a way of thinking. An argument often becomes clearer because you have been forced to put it into sentences, find the gaps and work out what you actually mean.

I would therefore keep essays.

I would also accept that much unsupervised learning will happen in an environment where AI exists.

Some students will write first and ask AI to criticise them. Some will use it to explain a difficult passage. Some will argue with it. Some will ask it for possible objections to their thesis. Others will use it for nearly everything.

Those choices will produce very different levels of learning.

I don't think education can completely control that from the outside.

The student who wants to learn will still have to decide how much cognitive work to hand over. The student who decides to hand over almost all of it may produce better-looking homework than before.

But then assessment needs to catch up.

We need to separate learning from verification

The job of education is larger than testing people. Reading, writing, discussing and experimenting all belong in the learning process.

The job of an assessment is narrower. At some point, we need to know what this particular person can actually do.

Those two functions have been bundled together.

An unsupervised essay can teach a student a great deal. It is becoming much weaker as stand-alone evidence that the student possesses the knowledge demonstrated on the page.

So leave more freedom around the learning process and introduce stronger points of verification.

A student could spend weeks researching and writing with whatever permitted tools they choose. Then ask them to explain the argument live. Give them an unfamiliar source and ask them to relate it to what they studied. Challenge one of their conclusions. Ask why they chose one interpretation over another.

The student who learned the subject, including the student who used AI intelligently

throughout, should be able to respond.

The student who outsourced most of the thinking has a much harder problem.

Suddenly we do not have to become detectives examining every suspicious phrase. Their understanding becomes visible.

This goes beyond universities

The same principle applies much earlier.

Rundell’s article is concerned with children and teenagers as much as university students, and here I think greater caution is justified.

A ten-year-old learning how to formulate an argument is in a different position from a postgraduate researcher who already possesses years of subject knowledge. If AI supplies the argument before the child has learned how to build one, the child may genuinely miss part of the development the exercise was designed to produce.

Children need periods when they read, calculate, remember, write and struggle without an AI system supplying the next step. Otherwise we risk testing their ability to operate a system before they have developed enough knowledge to judge what the system gives them.

But this still does not lead me to the conclusion that AI needs to be kept outside education altogether.

Children also need to learn what these systems are doing. They need to discover that a fluent answer can be wrong, that an invented citation can look convincing, that asking a better question changes the output and that the person using the system remains responsible for deciding whether the answer makes sense.

A child who has never been allowed near AI is not automatically prepared for a world containing AI.

A child who has never been required to think without it isn't prepared either.

The dissertation exposes the problem most clearly

Higher education eventually runs into a harder version of the same issue: the dissertation.

A dissertation can represent months of reading, reasoning and research. It can also carry enormous weight in determining a degree result.

And it is almost impossible to produce under supervision.

That makes it a particularly difficult assessment format in an era when generative AI can assist with research questions, structure, summaries, arguments, editing and large quantities of prose.

We cannot realistically put somebody in an examination hall for four months.

This is where assessment design may need a more serious rethink.

A dissertation or long research project could still exist, while the proof of individual understanding comes partly from something much shorter and harder to outsource:

• an oral defence in which the student has to justify methods, sources and conclusions;

• a supervised synthesis task based on unfamiliar material related to the research;

• a live application of the research to a new problem;

• a shorter written research project followed by a demanding viva;

• questioning that deliberately tests the weakest or most controversial parts of the submitted work.

A student could still spend months doing substantial research. The university would simply stop pretending that the final document alone is sufficient evidence of who did the thinking.

AI has made pretending easier. That is different from making learning obsolete

Rundell’s article is deliberately forceful. She describes AI as threatening the ability of young people to think, persist and develop intellectual independence, and ends with an argument for reading as a form of resistance.

I share much of her concern about what happens when a student mistakes generated fluency for knowledge.

I am less pessimistic about the student's desire to know.

The existence of a shortcut does not eliminate curiosity. It does, however, expose a weakness in systems that have relied heavily on completed work as evidence of learning.

Perhaps AI is forcing education to become more precise about something we should have been asking all along.

What are we trying to teach here?

What can the student use while learning it?

And when qualification, progression or a degree is at stake, how will we establish that they really know it?

Essays can continue to help people think. AI can become one of the things students learn to use and question. Independent reading and unaided thinking still need protected places.

And assessment has to become harder to fake.

That seems more realistic to me than trying to reconstruct a world in which students simply cannot ask a machine for the answer.

Reference: Katherine Rundell, “I hate what AI is doing to the minds and happiness of the young”: Katherine Rundell on the view from the classroom, The Guardian, 8 August 2026.Read the original Guardian article