What happens when you use AI as a servant, not a tutor

Almost every student now uses generative AI. The HEPI Student Generative AI Survey 2026 puts it at 95% using it in at least one way, and 94% using it for assessed work — up from 51% a year earlier. Savanta ran the fieldwork in December 2025 with 1,054 full-time UK undergraduates.

That first number gets quoted as a scandal. It is not one. Ninety-five per cent of students also use a search engine, and nobody writes policy notes about that.

The interesting question is not how many students use AI. It is what they are asking it to be.

Two students, same tool, opposite outcomes

HEPI's own report draws the line better than most commentary has: there is "a striking contrast between a student who is utilising AI to advance their learning and a student who is outsourcing their learning to AI."

Same tool. Same prompt box. Same 94%.

One student asks the model to explain why a step follows from the one before it, argues with the answer, and comes away able to reconstruct the argument without help. The other asks for the finished thing and hands it in. Both appear in the survey as a student who used AI for assessed work. Only one of them can do it again next week.

We call the second pattern using AI as a servant. You give an order, you receive an object, you never ask how the object came to exist.

The number people worry about is the wrong one

Twelve per cent of students now put AI-generated text directly into assessed work — up from 8% in 2025 and 3% in 2024. That figure gets the headlines, because it looks like cheating and cheating is easy to write about.

But copy-paste is the least interesting failure here. It is detectable, it is punishable, and the students doing it usually know exactly what they are doing.

The quieter problem sits inside the other 82%. These are students who rewrote everything in their own words, cited properly, broke no rule — and still could not tell you why the third paragraph has to come after the second. They outsourced something that leaves no trace in the final document.

What actually gets outsourced is the flow

A finished explanation has two parts. There is the content — the facts, the definitions, the formulas. And there is the flow — the order the parts have to go in, and the reason each step forces the next.

Ask a model for an explanation and it hands you both, fused together, already resolved. The content is visible on the page. The flow is not. It was decided somewhere inside the model, in a step you never watched.

You can read that output carefully, understand every sentence, and still absorb only the first half. Nothing warns you. The text makes sense on the way past, and making sense on the way past feels almost exactly like understanding.

This is why a student can produce good work in March and stall completely in June. The content was never the bottleneck. It was always available — in the textbook, in the lecture, in the model. What a degree actually tests is whether you can build the flow yourself when nobody hands it to you.

The gap stays invisible until something forces it open

Outsourced flow has a useful property, from the student's point of view: it is silent. There is no error message, no failed test, no moment where the shortfall announces itself.

It stays silent right up until one of four things happens.

A closed-book exam, where you have the facts but not the order. A viva or an interview, where someone asks "why that step?" and the answer was never yours. The next topic, which assumes you built the structure the last one was supposed to leave behind. Or a job, six months in, where the problem is slightly unlike anything you have seen and the shape of the solution is the only thing that would have helped.

By then the gap is months wide, and it is very hard to trace back to the semester where it opened.

Students are mostly doing this alone

Here is the part that shifts the blame away from students.

Read those four numbers together and the picture is unkind: nearly everyone is using it, the assessments have already moved, and roughly two-thirds are working out the rules by themselves.

If nobody shows you what good use looks like, the default is the fastest use. The fastest use is the servant pattern. That is not a character flaw — it is what happens when a tool arrives years ahead of the teaching about it.

What using it as a tutor looks like instead

The difference is not effort or time. It is a handful of habits, and each one forces the flow back into view.

Predict before you ask. Write down what you think the answer is, even badly, then ask. The gap between your guess and the answer is the actual lesson, and you can only see it if the guess exists first.

Interrupt. The moment a step does not follow, stop and ask why that step. Do not wait for the end. Reading to the end and then asking is how you end up with a second explanation you also did not follow.

Make it draw. Ask for the structure — boxes and arrows — before the prose. A diagram cannot hide the flow the way a paragraph can; if the arrows do not make sense, you find out immediately.

Explain it back, out loud, with the tab closed. This is the only one that reliably catches the illusion. Everything else can be done while half-convinced.

None of that requires a particular product. You can do all four in a plain chat window, today, for free.

Where this leaves us

We build HeyPapin around one of these habits specifically: it draws while it talks, and you can cut in mid-sentence. The interruption is not a convenience feature. It is the mechanism — you cannot interrupt an explanation you are not following, so the act of cutting in is proof that you were.

But we would rather be honest about the boundary. A tool can make the good habit easier; it cannot make it happen. A student determined to use any tutor as a servant will use ours as one too.

What we would take from the HEPI numbers is narrower and less dramatic than most of the commentary. The 95% is not the story. The story is that almost all of those students are deciding, without much guidance, whether the thing in front of them is a tutor or a vending machine — and that choice, repeated across a semester, is the whole difference.