AI in education: what to look for in the evidence

Schools and universities are adopting AI tools quickly, and the claims made for them range from modest to sweeping. Because learning is slow and hard to measure, the gap between a claim and the evidence is often wide.

What a convincing study looks like

A comparison group. The best studies compare students who used a tool with similar students who did not, ideally assigned at random. Without that, improvement could come from a new teacher, extra attention, or a different test.

A measure of learning, not speed. AI can help students finish homework faster. The question is whether they understand more afterwards, for example on a later test taken without the tool.

Enough students, for long enough. A pilot in one classroom over two weeks is a start, not a conclusion.

Independent authors. Studies run or paid for by the tool's maker are worth reading, but deserve a second look at how outcomes were chosen.

AI detectors

Tools that estimate whether text was written by AI are widely used, and also widely criticized. They can flag human writing as machine-written, and some research suggests non-native English writers are flagged more often. A detector result should be one signal among several, never the only evidence in a misconduct case.

Policy and guidance

International bodies such as UNESCO have issued guidance on generative AI in education, covering age limits, data protection, and teacher training. Guidance is not law. National and local rules vary.

How we cover education

We say who ran a study, who paid for it, how many students took part, and what was measured. We separate "students liked it" from "students learned more."