Every UK university that sponsors an international student has to know whether that student is still turning up. Home Office sponsor guidance requires a sponsoring institution to monitor engagement and to report a student who has stopped engaging with their course. Sponsorship can be withdrawn after that, and leave to remain curtailed. Nobody teaching on those courses asked for this arrangement. It is a condition of holding the licence that lets the university admit those students at all, and that is roughly the way it is.
Which means that sooner or later a university will ask us for engagement data. Not out of malice, and probably not from anybody who experiences the request as having a moral dimension at all. A compliance function asks its suppliers what numbers they hold, because that is what a compliance function with a legal duty does. Our answer is that we would not build it if a visa depended on it.
We should be honest about the reasoning we originally attached to that answer, because the reasoning was wrong. We told ourselves the risk did not really arise for us, on the grounds that we come in after the student already has an offer and that we are not the people marking the pre-sessional. Both statements are true. Neither of them is about the duty. The sponsor duty does not attach at admission. It attaches once a student is sponsored and studying, which is precisely the pre-sessional and in-sessional window our product sits in. And it runs through engagement rather than through attainment. Whether a student opened the app this week is the reportable kind of fact. How well they did is not. The refusal therefore applies exactly where we are, which is not where we first thought it applied.
Assessment would invert what the tool is for
The second refusal is assessment. If a university asked us to feed our output into a grade, we would say no, and we have two reasons for it that are not equally good.
The first is about compression. Putting a person's language into one number destroys the information that mattered. This is roughly the objection Krumm made to reporting a single CEFR level, in the 2007 exchange in the Modern Language Journal, in a piece titled "Profiles Instead of Levels". A learner who is B2 overall is not therefore B2 on vocabulary range and B2 on phonological control at the same time. Real students are uneven, and one letter is the representation guaranteed to hide it. A figure derived from app usage would be worse, because it would compress not only across skills but across a student's willingness to practise at all.
The second reason is better and it goes further than we would like. If our output counts toward a grade, the student's relationship with the tool inverts. They stop using it to learn and start using it to perform. Every mistake becomes expensive. What a language app is for, we think, is that you can be wrong in private at eleven at night in front of something that will not hold it against you later, and that property does not survive being marked.
We should be clear about the status of that sentence. It is a claim about what the thing is for. It is not a finding, we have no evidence for it, and since nobody has yet used our product for anything like a hundred hours there is no evidence we could honestly have. We also notice that the argument takes our own preferred compromise with it. We had thought a set of granular measures might be acceptable where a single grade was not. If the objection is really about performance rather than about compression, then granular measures that count are a more detailed way of making mistakes expensive.
Why we do not rewrite paragraphs
The third refusal is their coursework. Our writing feedback takes practice pieces only. It does not accept assessed work, and it will not rewrite a paragraph on request. A practice piece gets corrected where it needs correcting, and the student then has to write it again to get new feedback. We are aware that a more helpful-feeling product sits a small amount of engineering away from the one we have, and that we would sell more of it.
There are two reasons not to. The obvious one is that UK universities regulate third-party correction of assessed work through their proofreading policies, and some prohibit it outright. The less obvious one is that our own users are the population that gets accused.
The Office of the Independent Adjudicator published a casework note on 15 July 2025 finding that detection tools "may show bias against international students and those for whom English is not their first language, as well as disabled learners", and that the burden of proof sits with the institution. Liang, Yuksekgonul, Mao, Wu and Zou had already shown the same pattern experimentally in 2023, in Patterns, running writing by students who had learned English as an additional language through automated detectors. Building a tool that rewrites work for the group already most likely to be disbelieved would be a strange thing to do to them.
A policy is not a control
Now the part that matters more than any of the three, and it is the part a procurement process reaches quickly.
There is a difference between a policy and a control. A policy is what a company says it does. A control is a mechanism that makes the other thing impossible. "We tell students not to" cannot be entered into a risk assessment. Nor can "we would refuse if asked". Universities are not being cynical when they push on this. They have to write down what happens if the supplier's intentions change, or if the supplier is bought by somebody whose intentions were never the same. So it is worth stating plainly which of our three refusals are controls and which are currently only policies.
Refusing to rewrite a paragraph is close to a control. The system does not have that behaviour, and a student who asks for it does not get it. What we cannot promise is that no phrasing of the request ever gets nearer to one than we intended, because a system that generates language can be asked things in ways its builders did not anticipate.
Practice pieces only is mostly a control. There is no route by which assessed work enters the product: no connection to a submission system, nothing that ingests a marked assignment. What we cannot claim is that a text box knows what has been pasted into it. If a student pastes a paragraph of their real essay into a practice exercise, the system will treat it as practice, because from where it stands that is what it is. So the honest description is a control on the pathway and a policy on the content.
The refusals that are only policies
Grades are a policy. We do not produce an overall score for a university to export, and we have also not built anything that would stop us producing one later. Those are different facts, and only the second one is worth anything to a buyer.
The visa refusal is a policy too, and it is the one that should worry a buyer most, because the pressure behind it comes from a legal duty rather than from a preference. The control version is a written commitment, contractual and technical, that learning data does not reach admissions, recruitment or a compliance reporting function. We have not signed anything like it.
What the university is paying for
Which brings us to the thing we have been circling. Data about students is presently part of what we offer a university. That is not a temptation we are describing in the abstract in order to look clear-eyed about it. It is in our own material: the university pays, and part of what it pays for is data collection for the university's benefit. One of the people advising us put it to us that the app could be used to infer things about students who might come to a given university, and that it might be used to make offer-holders more likely to accept. We took that seriously enough to write it down.
The objection to that arrangement is older than the software. Joan Turner made it against pre-sessional courses in 2004, and Alex Ding and Ian Campion gave it wider circulation when they quoted her in 2016: a quick fix attitude persists in pre-sessional provision, and it perpetuates a philosophy of "maximum throughput of students with minimum attainment levels in the language in the shortest possible time". Turner was describing teaching rather than software, and the extension is ours rather than hers. But a learning product whose business case is offer-holder conversion does not sit next to that objection. It is that objection with software attached, one step earlier in the process, before a student has accepted a place.
The legal position is not much friendlier. A student consents to a tool that helps them learn academic English. Using the behavioural record that tool generates to score their likelihood of accepting an offer is a materially different purpose, and under Article 5(1)(b) of the UK GDPR consent for the first does not carry over to the second. If the university is the controller and we are the processor, then generating new inferences the university never instructed sits outside a processor role entirely. We have been advised that data protection is a solvable assessment step, and for the learning product we think that is right. The profiling proposition is a different kind of problem. It is a design decision, and design decisions get made by people rather than by lawyers.
Two versions of this company
We have not made that decision. There is a version of this company that signs the firewall, gives up a line of value it has already put in front of investors, and can tell an EAP centre that nothing a student does inside the app will ever be used to sell to them or to report on them. There is another version that keeps the option open because it is worth money and nobody has yet forced the question. We know which one belongs on a page like this. We have not done it, and an essay that claimed otherwise would be worth nothing to the people reading it, since the claim would be checkable inside a year.
The test of this is what gets built, and which of the two pitches wins on the day they actually conflict. A university that puts its name to a supplier takes on that supplier's positioning and not only its feature list, and the bar we set ourselves was a product a university would be proud to have made itself. On the philosophy of it we do not currently clear that bar. Whether a company can hold a firewall in place while the thing on the other side of it is still sitting in the business plan is not something we can settle by writing, and we would rather say so than pretend the decision has already gone the right way.
