Your Students Are Already Using AI You Cannot See

The debate at most institutions is still framed as whether to introduce AI into student writing. That decision was made without you, roughly two years ago, by the students.

In HEPI’s Student Generative AI Survey 2026 — Savanta, December 2025, 1,054 full-time UK undergraduates — 95% reported using AI in at least one way and 94% said they use generative AI to help with assessed work. In the same survey, 38% said their institution provides them with AI tools.

The distance between those numbers is the actual operational problem. It is not a compliance gap. It is a provisioning gap, and everything uncomfortable about the current situation follows from it.

What the gap actually costs you

Every student in that gap is doing academic work on a consumer account. Four consequences follow, and none of them are hypothetical.

You have no agreement covering the data. Postgraduate work in particular carries unpublished research, sometimes interview transcripts collected under an ethics approval that named no commercial processor. Your institution has a data protection position for every system it procures and none at all for the tools most of your students are actually using.

You have no visibility into how work was produced. When a supervisor or a panel asks how a chapter came to exist, there is nothing to look at. That is the same absence that makes integrity cases turn on inference — the problem we set out in our analysis of AI detection reliability.

You have an equity problem you did not choose. Capability now tracks what a student can pay for privately. When 68% believe AI skills are essential and 38% are provided with tools, the difference is being funded out of household budgets, and it shows up in work quality as something that looks like ability.

Your policy is not reaching people. In the same survey, 37% of students agreed their institution encourages AI use and 36% disagreed. A near-even split at the same institutions means the rules are not landing, whatever the rules say.

What this looks like from the supervisor’s side

The institutional framing above is the one that reaches a committee. The version that reaches your academic staff is smaller and more corrosive.

A supervisor receives a chapter that is noticeably better written than the student’s previous work. Nothing about it is provably wrong. There is no source to check, because the concern is about authorship rather than copying. Raising it means initiating a process against a student who may well have simply improved; not raising it means a quiet erosion of confidence in the supervisory relationship. So in most cases nothing is raised, and the doubt is carried privately.

That is the real cost of having no process evidence, and it does not appear in any caseload statistic — because the whole point is that these cases never become cases. Ask three supervisors in a research-intensive department whether they have had that experience in the last year. The answer will tell you more about your institution’s exposure than your integrity numbers will.

Why a stricter rule does not close it

The instinct is to tighten the policy. It is worth being clear about why that does not work on its own.

A prohibition you cannot enforce transfers the behaviour from visible to hidden without reducing it. It also penalises exactly the wrong students: the conscientious ones who read the policy and comply, while their peers proceed unaffected. And it puts the institution in the position of forbidding something a growing share of its own curriculum now teaches.

What does work is narrower and less satisfying: decide what is permitted per assessment, publish it where students will actually see it, provide the tool where you expect it to be used, and require disclosure. The components are set out in what a university AI policy should include. Provision is the component that turns a policy from a statement into an operating reality.

What Tesify for Institutions changes

We are a writing platform, not a detection product, and the distinction matters for what we can honestly claim.

• Work happens inside an institutional environment. One agreement, one data protection position, one set of terms — instead of an unknown number of consumer accounts you have no relationship with.
• The writing process leaves a record. Structure, drafts and revisions exist as artefacts, so a supervisor or panel has something inspectable rather than an inference about a finished document.
• Support scales past staffing. Structural and citation help is available to every student at once, which is not a thing a timetabled service can do.
• Access is equal by construction. Provisioning at institutional level removes the private-purchase gradient from your cohort.

What we will not tell you: that this identifies misconduct, or that it makes detection unnecessary. It does neither. It changes what evidence exists about how work was produced, which is a narrower and more defensible claim — and it is the claim the incumbent’s own product positioning has now moved toward.

Nor will we ask you to displace a screening contract. Screening examines work after submission; a writing platform changes how work is produced. They answer different questions, and any business case pretending otherwise will not survive its second year.

Start with a departmental pilot, not a purchase

The reason to start with a free departmental pilot is procedural rather than commercial: a pilot clears the internal path that a purchase order does not. It produces evidence against criteria your own committee agreed in advance, and it does so without committing budget while the evidence is still absent.

A pilot that answers the question needs four things, all of which we will help you put in place:

1. A baseline captured before launch — writing centre demand, supervision time, integrity caseload, and a short student survey you will repeat verbatim.
2. Success and stop criteria signed before anyone logs in.
3. The data protection review running in parallel, not afterwards.
4. A decision paper written against the criteria, not against impressions.

The full structure, with owners and artefacts for each step, is in our guide to running a departmental pilot. We will supply the pilot brief, the criteria template and the data protection documentation as a package so your team is not building them from scratch.

One scheduling note worth acting on early: because a credible pilot needs a baseline captured in the preceding term, the practical lead time is longer than the pilot itself. An institution that starts the conversation at the beginning of a term can usually pilot in the next one; an institution that waits until a problem is acute will find that the evidence it needs had to be collected before the problem became acute.

More than 9,000 students already write with Tesify and more than 15,000 chapters have been drafted on it, with the argument and analysis remaining 100% written by the student throughout. The institutional version puts that inside your governance rather than outside it.

Request an institutional evaluation and we will start with your problem statement rather than a demonstration.

Frequently asked questions

What does a pilot cost?

A departmental pilot is free. It is scoped to one department and one term so it produces a real result rather than an impression.

How long does a pilot take?

One academic term, plus preparation in the term before for baseline capture and the data protection review.

Do we have to replace our existing detection contract?

No, and we would advise against changing two variables at once. Screening and writing support are complements.

Where is data stored and is it used to train models?

These are exactly the questions to put in writing before any call. Our procurement question bank lists them, and we answer them in writing for our own product.

How does this fit with GDPR and FERPA?

Through a data processing agreement, a named sub-processor list and retention terms. The review sequence is in our data protection guide; note that in US higher education the FERPA rights holder is the student.

Does this help with academic integrity?

Indirectly and honestly: it produces process evidence and supports disclosure. It does not adjudicate misconduct, and we will not claim that it does.

What integration is required for a pilot?

Usually less than for a production deployment. Keep them distinct so a smooth pilot is not mistaken for evidence about full integration.

Who should sponsor the pilot?

An academic owner with authority over assessment — a dean, graduate school director or writing centre lead — with IT and the data protection officer supporting.

What if the pilot shows no benefit?

Then the criteria sheet says stop, and you have saved a procurement cycle. A pilot that cannot fail is not worth running.

Can we run it with postgraduate researchers only?

Yes, and it is often the clearest population to start with, because thesis work has visible milestones to measure against.

How do we justify this to a finance committee?

On staff time and caseload against your own baseline, not on national statistics. Why the national figures cannot carry that argument is explained in what student AI statistics actually measure.