AI in Schools: What Actually Works, and What Is Still Marketing

A Grade 8 science teacher is handed a new AI tool in a Monday staff meeting. The demonstration runs fifteen minutes, the phrase “saves hours every week” is used twice, and there is a login sheet by the door. By Thursday the tool has been opened by four people out of thirty, three of whom were presenting it. Nobody is being obstructive. The tool simply did not touch anything that was actually on those teachers’ desks that week.

That pattern is the most reliable thing about AI in schools right now, and it is a useful filter. The tools that survive a term do one narrow thing: they take work an adult is already doing slowly, do a passable first version of it, and hand it back in a form the adult can still argue with. The tools that die ask a teacher to change how they teach in exchange for a benefit that arrives, if at all, next year.

Three things that genuinely work

Drafting the material nobody enjoys making

Producing twenty questions on the water cycle at three difficulty levels is not intellectually demanding work. It is slow, repetitive, and it happens on a Sunday evening. A model does it in under a minute, and the quality bar it has to clear is not “better than a good teacher with unlimited time” — it is “better than the question bank that has not been touched in four years”. That is a bar it clears easily.

Two conditions decide whether this actually helps. The output has to be editable and rejectable, so authorship stays with the teacher rather than being quietly transferred to a vendor. And the questions have to be tagged with the concept each one tests, because that tagging is what makes the results diagnostic later instead of just a column of marks. If you are evaluating something in this space, how the assistant hands work back to the teacher matters more than how impressive the generation demo looks.

Explaining the same thing for the fifth time, at nine in the evening

The value here is not intelligence. It is patience and availability. A student whose parent can explain simultaneous equations loses nothing when the school day ends. A student whose parent cannot loses the entire evening. That gap is one of the largest and least discussed differences between two children in the same classroom, and it is one of the few things software is genuinely well shaped to reduce.

The design constraints are strict, though. It has to explain rather than supply finished answers, it has to work from the material the school actually teaches rather than whatever the open internet contains, and the session has to leave a record the teacher can look at. A tutor that quietly does the homework is worse than no tutor, because it also removes the signal the teacher would have used to spot the problem. Those constraints are the whole design question behind an AI tutor that is worth putting in front of children.

Turning results into something you can act on

The distance between “37 per cent” and “she cannot subtract fractions with unlike denominators” is the distance between a report and an instruction. Closing it is mechanical work: map items to concepts, group students who failed the same concept, order those groups by how much downstream content depends on the broken idea. It is exactly the sort of boring, high-volume bookkeeping that should be automated, and it is far more valuable to a school than anything generative. This is the part of an AI platform for schools that changes what happens on Monday.

Three things that are still mostly marketing

“Personalized learning” that is a branching quiz

Personalization has a testable meaning: the next thing a student does is chosen because of what they specifically got wrong. Much of what is sold under the word is a two-branch decision tree sitting on top of a video library, which serves easier content to students who score badly and harder content to students who score well. That is streaming with a shorter feedback loop.

There is a question that separates the two quickly. Ask what the system does for a Grade 9 student whose real problem is a Grade 5 gap. If the answer involves more Grade 9 practice, delivered more slowly, the system cannot reach the actual break. A path worth the name goes backwards to the prerequisite and then returns.

Engagement and attention scoring

Camera-based attention detection and time-on-page metrics measure compliance and household bandwidth. A student reading a paragraph three times because it is difficult looks identical to a student who has wandered off, and a student who understood it immediately looks disengaged. Beyond the measurement problem there is a behavioural one: once children know attention is scored, they perform attention. And the privacy cost of continuous observation of minors is not a footnote — it is the main term in the equation.

Risk scores with nothing behind them

A flag that says a student is at risk without saying which evidence produced it cannot be acted on, cannot be challenged by the teacher who knows the child, and cannot be corrected when it is wrong. It also has a habit of hardening into a label that follows a student between teachers. A risk indicator is only defensible when the reasons sit one click away and the recommended action is specific, which is the standard worth holding any early-warning system to.

How to tell the difference before you sign

Most of the useful diligence questions are unglamorous:

• Which specific task disappears from someone’s week, and whose week is it?
• What happens when the model is wrong? Who sees the error first, and what does it cost to fix?
• Is the content mapped to the syllabus you teach, or to a generic library?
• Can a teacher see the evidence behind every number shown to them?
• What does the school keep if the contract ends — the content, the assessment data, the concept mapping?

Then run it on one real class for one term, with one measure agreed in advance. Teacher hours spent on preparation is a good measure because it is honest and easy to collect. Re-assessment on the same concepts eight weeks later is better, because it is the only one that speaks to learning rather than convenience.

The part no tool fixes

Adoption follows workload relief. A school that adds a platform without removing anything has issued an instruction, not made a change, and instructions produce logins rather than use. The tools that stick are the ones a teacher would open on a Thursday without being asked, because the alternative is doing the same task by hand.

None of the cases that work require believing anything dramatic about artificial intelligence. They require being specific about which parts of a teacher’s week are worth machine effort, and honest about which parts are not. The second list is longer than most vendors admit, and being clear about it is the fastest way to find the tools that will still be in use in March.