AI in school: what the evidence actually says
Two results from the past year point in opposite directions — and the difference between them is the whole of what schools need to teach.
The evidence does not say AI helps or harms students. It says the same tool does both, and which one you get depends on whether the student is still doing the thinking. Two results from the past year make the point better than any argument about policy:
An AI tutor outperformed in-class active learning by 0.73 to 1.3 standard deviations — a very large effect by the standards of education research.
Frequency of AI use correlated negatively with critical thinking. The mediation analysis named the mechanism: cognitive offloading.
Same technology. The students who gained were using it to get explanations and then doing the work. The students who lost were using it to get the answer. The trial and the review of offloading harms are not in conflict; they are describing two different behaviours that look identical from the doorway.
Schools are already being told to teach it
This has moved from debate to deadline faster than most staffrooms have noticed. As of mid-2026, 37 states and Puerto Rico have issued official K-12 AI guidance, and 27 states have live legislation. Maryland's Ready Schools Act requires AI literacy to be built into state standards by 1 June 2027.
What to teach is less of an open question than it looks, because UNESCO has published a competency framework and it is unusually sane. Four strands:
The student understands they have agency here — that a system's output is a proposal, not a verdict.
Responsible use, safe practice, and the idea that design choices carry values.
What the thing actually is. Enough mechanism to reason about it.
Problem-solving and design thinking — building, not just consuming.
Notice what is missing: a list of tools. The framework is about understanding, and it is meant to run through existing subjects rather than beside them as a bolt-on unit. That is the part most schools get backwards.
The detector problem, which is worse than most schools know
The first instinct on hearing that students use AI is to buy detection. It is the one decision here with hard evidence against it.
Researchers ran seven AI detectors over 91 TOEFL essays written by non-native English speakers and 88 essays by native speakers. The native-speaker essays were classified near-perfectly. Of the non-native essays, 61.3% were falsely flagged as AI, and 97.8% were flagged by at least one detector.
The cause is not malice, it is arithmetic. Most detectors score text on how predictable it is, and someone writing in a second language reasonably reaches for simpler structure and more common vocabulary. That reads as machine-like to a perplexity score.
“The detector said 92%” is the first scenario in our own 5-minute lesson for teachers, and it is there because a percentage is uniquely persuasive. It is not an accusation a teacher can put to a student, because the teacher cannot explain how it was produced.
What using it well actually looks like
The offloading research points at a practical line, and it is finer than “allowed” or “banned”. The gains came when AI carried the explanation and the student carried the reasoning. The losses came when it carried both.
In classroom terms that is the difference between “explain why this step works” and “do question four”. The first leaves the student with something; the second leaves them with a page.
There is a second skill, and it is the one this site keeps returning to: an AI does not know when it is guessing. A fact and an invention leave it in the same confident voice. A student who cannot tell which is which has not been given a tool, they have been given a plausible-sounding stranger — which is exactly what why AI makes things up is about, and what the app's Know It or Guess It? level drills directly.
What teachers should not hand over
Three of these come up constantly, and all three are places where the tool is genuinely tempting:
Grading at scale. It will produce a mark for every script tonight. It cannot tell you which child has quietly stopped understanding, and that was the actual point of marking.
Lesson plans. The plan will look excellent, because it is built from plans that were excellent somewhere else. It does not know about the two students at the back, the trip on Thursday, or what this class already found hard.
Student records. “More detail gives a better answer” is true and it is also how identifiable information about a child ends up in a chat box on someone else's servers. The question to ask is not whether the output improves — it does — but whether you had the right to put that in.
And the career question
Parents ask what to steer children toward, and the honest answer is that nobody knows. A child in Year 7 today reaches the labour market around 2034, and confident predictions about which jobs survive are worth exactly what they cost to make.
What is defensible is narrower and more useful. Every one of those futures contains work where a machine produces something and a person decides whether it is any good. The transferable skill is judging the output — noticing what it assumed, what it invented, and who it left out. That is not a subject on a timetable. It is the same habit the whole of this article is about, and it is worth more than a list of safe careers.
Read next
Was this useful?
Would you want your child to know this?
FutureMinds teaches 9–16s to tell what an AI knows from what it is guessing — by playing, not by being warned. Take the 2-minute AI Smart Challenge yourself first — five real questions from the app, no signup.
Take the Challenge →
Join the conversation
Comments are read by a person before they appear — usually within a day or two. Nothing is published automatically.