Control over how AI is used in a classroom is rarely taken from anyone. It gets handed over quietly, one convenient and small step at a time. Understanding what a tool is built to do — and what that design means — is how the decisions stay yours.
Ask a teacher whether a textbook is any good and you will rarely get a straight yes or no. Most probably you will get a question back. Good for what? For which year group? Alongside what else?
That instinct — refusing to judge a resource without knowing what it is for — is the most useful thing anyone can bring to generative AI. It is also what keeps a person in charge of the automation rather than the other way around. Nobody hands over a professional decision deliberately. It goes by default, in small pieces: a suggestion accepted because it appeared first, a plan used because rewriting it would take longer, an answer trusted because it was phrased confidently.
They are not one thing
Public conversation treats AI as a single product, so the question becomes whether “AI” is good or bad for learning. This question has no answer; because it is the wrong one!
When our taxonomy for GenAI4Ed work began, the first task was not assessment but description: what is each of these tools actually built to do? The answers were not close together. A lessonplanning assistant, a conversational tutor, a writing companion and an image generator have about as much in common as a microscope and a telescope. Comparing them on general merit is a category error — the only sensible question is which one suits the thing you are trying to see.
Strengths and pitfalls are part of the same decision
Each tool was designed by people with a particular job in mind. That decision explains what it is unusually good at. It also explains, quite reliably, where it will let you down.
- A tool built to produce fluent, confident prose will produce it.
Including when the content is wrong. The polish is the feature; it is also what hides the error.
A tool built to answer immediately will answer immediately.
Exactly what a stuck student needs at nine in the evening — and exactly what removes the productive struggle where the learning happens.
A tool built to adapt to a learner needs to know things about that learner.
Personalisation and data collection are one mechanism, described from two directions.
None of these is a defect somebody forgot to fix. Each is the cost of the tool working exactly as intended.
A tool’s pitfalls usually sit right beside its strengths — the same design decision, seen from the other side.
Understanding is what keeps the decision yours
None of this requires auditing a model. It requires knowing what the tool is for and asking what follows from that — a question a teacher can ask in a free period and a parent can ask at the kitchen table.
Once it has been asked, the decision moves back to the person. If you understand that a tool sounds equally certain whether or not it is right, you are the one who decides where that is acceptable and where it is not. Without that understanding, the same decisions still get made. They are simply made by whoever designed the default.
THREE QUESTIONS THAT KEEP THE DECISION YOURS
- What is it built to do? Not what the marketing claims — what the tool is clearly optimised for.
- What does that strength cost? Speed, fluency and personalisation each come with something attached.
- Does that trade suit this learner and this task? The same trade can be right for a sixth-former revising and wrong for a nine-year-old practising.
Where our work fits
Much of what GenAI4ED is building comes down to carrying this knowledge so that people do not have to derive it every time: describing what each tool is for and what comes with it, so that recommendations arrive with the trade-off attached rather than as a bare ranking. What the platform will not do is decide for you. It cannot know your class, your school’s policy, or the student who has had a difficult term. The aim is not to automate professional judgment but to make it cheaper to exercise.
Confidence, not caution
It would be easy to read this as a warning. But it is not meant to be; exactly the opposite. Knowing where an instrument stops is not scepticism about the instrument; it is what fluency with any tool looks like. Understanding does not slow people down — it is what lets them move faster without wondering what they have given away.
Bringing AI into education will be a long process, with a fair amount of revision in it. It is worth being clear from the outset that this is a journey we are steering, not one we are being taken on. The tools are built for particular jobs. Tools built by Us, humans, for our own benefit. Deciding what to do with them is our responsibility.
Author: Petros Gikas