Ten Positions on AI in School

In the loop is a phrase most suppliers use, and what it usually amounts to is that the teacher is shown a conclusion and given a button. Authority is not a button. It is the answer to three questions a school can put to any supplier: what does it cost a teacher to disagree, what is her disagreement then treated as, and what becomes of her decision afterwards.

The first is friction. If accepting a recommendation takes one tap and rejecting it requires a typed justification, the design has put a thumb on the scale, and over a term of fifty-five-student sections we would expect the thumb to win. What that produces is not agreement but deference by fatigue, and in every record the school keeps the two look identical. The second is interpretation: a teacher’s disagreement is not a compliance failure, and a school is entitled to ask where it goes and who is shown it. The third is persistence: a decision the system quietly reverses later was never the teacher’s to make.

The objection we take most seriously is that teacher judgement is itself patterned, and that the patterns track language, gender and class. If that is so, it is an argument for putting the contradicting evidence in front of her in a form she can examine and reject. It is not an argument for moving the decision away from the only adult who knows the child, sees the home situation, and will be standing in front of the parent in April. Responsibility without authority is the fastest way to lose a good teacher’s engagement entirely.

Two things could be meant by explanation and only one matters here. A parent does not need an account of how a system works. A parent needs the grounds: what the child actually did, when, and what was read from it. That distinction is what makes the requirement achievable without publishing a mechanism, and also what makes it demanding, because grounds can be checked and mechanisms cannot.

An explanation that cannot be checked is a reassurance. Telling a family their daughter is a visual learner, or that she is at risk, gives them nothing to inspect and nothing to contest. Telling them what their daughter actually did, and what was read from it, gives them something they can argue with. The second statement can be wrong in a way a parent can spot. That is the property worth having.

Language is not a presentation layer here. The parent may not read English, may not have finished school, and may be reading on a shared phone at the end of a working day. An explanation that exists only in the medium of instruction puts the child in the position of translating an assessment of herself to her own parents, which is unreliable and unfair to her. If an inference cannot survive being stated plainly in the home language, it was never clear enough to act on. And where the grounds are thin, the honest move is to make no claim at all. Saying nothing about a named child is a cost a school can weigh; saying something that cannot be shown to her parent is not.

A system that produces well-formed prose is optimised, by construction, for the appearance of knowing. It produces the register of the textbook, the shape of a derivation, the cadence of an explanation that has been checked. None of that is evidence about the content. This matters more in a classroom than almost anywhere else, because the recipient is a novice, and a novice is defined by being unable to tell a correct explanation from a fluent one. A student who could catch the error would not have needed the help.

The cost is asymmetric and usually described too weakly. A wrong answer does not merely fail to teach; it teaches. It installs a misconception with the authority of the school behind it, at the moment the student is most receptive, and that is harder to remove than the ignorance it replaced. Underneath sits a worse lesson: that confidence and polish are the marks of truth. Examinations punish that habit, and so does everything after them.

Refusal therefore has to be an ordinary, visible, unembarrassed state, and a system used with children should decline more often than one used with adults. Refusal reduces usage, and usage is what products are measured on, including ours. We take usage to be the wrong objective: a tool used less and trusted properly is better than one used constantly and believed uncritically.

The crude version of this position is that students should not have machines do their homework. It is too crude to use. The pupil who copies a solution has learnt nothing; the pupil who spends the hour on arithmetic slips during a lesson about forces has also learnt nothing about forces. Every task contains effort that builds the capability being taught and effort that is only the price of entry, and removing the second is one of the few unambiguously good things a machine can do in a classroom.

The difficulty is that the boundary is not a property of the task. It belongs to the pair: this task, this child, this week. Rewriting a messy solution neatly is the entire learning for a student who cannot yet organise an argument, and pure transcription for the girl beside her who organised it in her head twenty minutes ago. Look up a formula and you have either removed a distraction or removed the retrieval practice that was the point.

So we state this as an open problem rather than a policy. We do not know how to locate the load-bearing effort in a task reliably, for an individual child, across subjects, and we are not close. Anyone claiming to have solved it should be asked what result would have shown them wrong. What follows for a school is modest and firm: a line drawn once, for everyone, and left unstated is the wrong shape, and the people entitled to move it are the teacher and the child in front of her.

We have not conducted a systematic review of the evidence on AI and learning, and this page is not one. What can be said without a review is structural: any result a supplier shows a school was obtained in some other classroom, because it was not obtained in this one. That is not an accusation, it is arithmetic, and it puts transferability on the buyer rather than on the seller. A result from a class of twenty may hold in a section of fifty-five sitting boards in a second language on shared devices, with a syllabus that must be finished by February. It may not. Nothing in the result itself settles which, and the school is the only party in the room with an interest in finding out.

Four questions cost a school nothing to ask. On whom was this measured, described precisely enough to compare with our children. Compared with what, meaning what the other group received rather than whether one existed. Who paid, and who chose what was reported. And the question that separates research from marketing: what result would have counted as failure. If no outcome would have been reported as a failure, nothing was tested, and what is on offer is a hypothesis with a price.

The standard applies to us, and by that standard we have published nothing. This site carries positions and questions. Where a reader expects a number there is none, and the omission is deliberate rather than coy: a figure detached from its population, its comparison and its failures is a decoration, and an accurate decoration is still not evidence. When work here reaches a point where population, comparison and negative findings can be published together, that is the form it will take.

The ordinary situation in an Indian classroom is that instruction happens in one language, thinking happens in another, and the explanation that reaches the parent happens in a third. Children reason in the language they grew up in and write in the language of the examination, and there is loss in that passage. The loss is a translation gap and it is systematically mistaken for a knowledge gap.

That mistake is a mechanism by which capable children are sorted downwards. A system reading only what a student wrote in English records a vocabulary failure as a physics failure, and if that record then decides what she is given next, she is moved into easier material for a reason that has nothing to do with physics. Two states look identical on paper and need opposite lessons: the child who does not know the concept, and the child who knows it and cannot yet say it in the register the examination demands.

Translation does not settle this. Rendering a question into another language changes the question, because subject vocabulary often has no settled equivalent and because examination language is a dialect that has to be learnt eventually; removing it from a child’s path is deferral, not kindness. Mixed-language work, a sentence begun in one language and finished in another, is ordinary speech and not an error. A system that treats it as a mistake is telling a child that the way she thinks is a defect.

An improvement in the class average is entirely compatible with harm. If a tool helps the confident students and confuses the students already behind, the mean rises, the spread widens, and the summary reports success. This is the expected failure mode rather than a rare one, because almost every tool asks something of its user, and the students best able to supply it are the ones who needed it least.

The lowest group is hardest to serve for structural rather than motivational reasons. They have the least prior knowledge with which to catch a wrong answer, so fluent errors land on them hardest. They read slowest, so interfaces built for comfortable reading exclude them first. They have the least reliable access to a device, so anything requiring continuity outside school fails for them specifically. And they complain least, so their difficulty appears as quiet non-use rather than as a defect report.

The position is that a tool is judged on whether the bottom of the room moved, and that a tool raising the mean while the bottom falls further behind should be called a failure. Against this it can be argued that total learning is what matters, and that a large gain concentrated among the students able to use it is still a large gain. A school is not a portfolio, though. It exists in large part for the children who would not get there otherwise, and those children remain in that room for years afterwards.

The compliance question asks what a school can be brought to consent to. The design question is harder and more useful: what is the smallest amount of information about this child sufficient for the decision being made now, and how soon can the rest be released. The first produces a consent form. The second produces a different system.

Children’s records are not ordinary personal data. The person described is a minor, the consent was given by someone else, and the record outlives both the child’s time at the school and the school’s relationship with the supplier. Something written about a thirteen-year-old having a bad term should not still be a fact about her at twenty-five. She had no practical means of contesting it when it was written and will have none later, because by then it will be a number in a summary whose origin nobody remembers.

The cost is real and we accept it. A broader record about a child would probably support better inferences, and holding less means accepting worse ones. We take that trade for children, and the argument against it is a serious one: that a more accurate system helps a child more than a leaner one protects her. It should be had in those terms rather than settled by a checkbox.

The case against the mark is set out on the position page and is not restated here. What this position needs from it is narrow. A mark is a reasonable instrument for ranking and a poor one for teaching: it identifies who is behind, and it does not tell a teacher what to do with the forty minutes she has tomorrow, in front of fifty-five children, with a syllabus that must be finished on time.

The useful output of an assessment is therefore an action rather than a score. Reporting that a chapter went badly is not a lesson plan, because a chapter is not something anyone can teach in a period. Reporting a specific misunderstanding that a teacher could take up in one period is. Timing belongs to the instrument and not to operations: an assessment arriving after a unit has been taught cannot change how it was taught, only inform a repetition the calendar does not allow.

Two arguments run alongside this position rather than against it. The first is that a class taught only what it got wrong receives a fragmented subject, assembled from its own errors and with no shape. Diagnosis is an input to planning and not a replacement for it, and a teacher who chases a list of misconceptions will produce students who can patch errors and cannot see the whole.

The second is the argument this site makes against the mark, turned on whatever is proposed in its place. A measurement that compresses deforms the curriculum, and it does not follow that a measurement which compresses less is safe. Any account of a learner more detailed than a mark, used inside a system that allocates scarce seats, offers more surfaces to teach to rather than fewer, and more places where a record can be curated before anyone reads it. It may deform teaching worse than the mark does, precisely because it is more specific about what to deform. We do not know how to prevent that. Three things would tell us it had started: a school reporting such an account upwards instead of teaching from it, a teacher appraised on what it says about her section, and a parent who asks to have it improved rather than asking whether it is true.

This is not a call for general caution. It is a list, and a school can adopt it without buying anything or waiting for anyone’s research.

Irreversible sorting comes first: streaming, gatekeeping, selection, exclusion, and any prediction of how far a child will go, which we expect to arrange its own confirmation, because the adults who see it teach accordingly and the child eventually hears it. Second, inference about a child’s inner life: emotional state, motivation, character, family circumstance, mental health. Defensible across a population, such an inference becomes, when attached to one named child, an unfalsifiable claim a school will nonetheless act on. Third, discipline. Fourth, unsupervised communication with a child, which schools already know how to think about and should not stop thinking about because the counterpart is a system. Fifth, public comparison of children against one another, a bad practice before any of this and not improved by automation.

The first item carries a real cost. Identifying children at risk of dropping out is how a school helps them, and refusing the prediction has a cost of its own, paid by children nobody noticed in time. The same flag that marks a child for support also marks her for lowered expectations, and in a school under pressure, with too many children and too few hours, we would expect the second reading to be taken more often than the first. The objection is therefore not to the prediction but to the institution receiving it. Where a school can show that a flag reliably produces extra teaching rather than reduced ambition, the position should be revisited for that school.