AI helps students most when they are taught to challenge it

AI is not a magic tutor, according to a new analysis of OECD education data. The reported finding is more interesting than the usual “students use chatbot, civilisation ends” routine: outcomes depend heavily on what students ask AI to do and whether they are taught to question its answers.

Watch Desk analysis

What happened

The Verge says the study analysed 2025 data from more than 760,000 15-year-olds in 91 countries, covering science, maths and reading. After adjusting for socioeconomic status, students who never used AI generally outperformed those who did in science.

The relationship was not uniform. Students using AI to summarise texts they were meant to read or draft assignments performed relatively worse, while those using it for preliminary research or general help with learning saw a smaller drop. Students using AI weekly for general learning slightly outperformed non-users, and the advantage grew among regular users who were frequently asked to assess the quality of AI-generated information.

Key findings

• The broad result
  Students who used AI generally performed worse than non-users after socioeconomic adjustment, particularly in science.
• The task matters
  Summarising assigned reading and drafting work were associated with worse performance than using AI for preliminary research or general learning help.
• Frequency is not a straight line
  Daily users and people using AI only once or twice a year performed worst; monthly and weekly users did better in the reported comparison.
• Critical assessment changes the picture
  Regular users who were taught to evaluate AI-generated information outperformed non-users when using AI for general learning.
• Access is uneven
  Reported AI use ranged from more than 95% of students in Vietnam to about 60% in Japan, reflecting differences in access and context.

Why it matters

The useful lesson is not “ban the chatbot” or “give every pupil a prompt subscription”. It is that AI appears to work better as part of an active learning process than as a vending machine for finished answers.

For students, that means using AI to investigate, test and critique ideas is more promising than handing over the reading or the thinking. For schools, teaching pupils how to check AI output may matter as much as deciding which tools they are allowed to use.

Our read

This is a timely warning against measuring AI in classrooms by adoption alone. The strongest reported result points towards a practical rule: preserve the productive struggle, and make evaluation part of the lesson. The underlying OECD publication still deserves a careful read before anyone turns these associations into a universal classroom prescription.

What to watch

• Whether the OECD publishes fuller task-level results and methodology.
• How much of the observed relationship reflects prior attainment, access, motivation or school policy.
• Whether schools that explicitly teach AI evaluation reproduce the reported improvement.
• Whether education systems respond with targeted guidance rather than blanket bans.

Discussion spark: Should schools teach students to use AI as a tool for investigation and criticism, or is that still too dependent on learners already knowing what good work looks like?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.