The students are ahead of us

The episode with Ryan Trattner went live on Sunday, one of the few founders building in the AI tutoring space who is willing to sit with the harder version of the question rather than the marketing version. The blog companion followed on Monday: The question has changed, from policing AI to rethinking with it. It is the version of the argument I would want a school leader to read before their next staff meeting.

The bit that did not make the blog: I asked Ryan whether he worries that the better his product becomes, the less his users have to think. He did not dodge it. The moment a tool becomes frictionless, he said, is the moment it stops being a teacher. That is not the kind of sentence you usually hear from a founder talking about their own product.

Published 9 June. The most useful piece of research the UK secondary sector has had in months, and the one that has shifted my thinking most this week. Nearly four thousand young people surveyed; three thousand one hundred aged thirteen to eighteen; more than seven hundred in the qualitative work.

Three numbers I keep returning to. Fifteen per cent say school has given them enough guidance. Forty-four per cent think using AI for all of their homework is cheating. Seventy-three per cent say AI cannot replace a teacher’s empathy.

This is not the panic narrative. It is a population of teenagers thinking carefully about a tool they do not yet have a settled relationship with, asking the adults around them for guidance they are not getting. The guidance gap is the actionable finding. Most schools are spending their time writing policies students do not find useful. Students are asking for something different: how to prompt, how to check outputs, and when not to use the tool at all.

Several projects are moving into beta over the next few weeks. First to open is the Lead with AI courses. There are a handful of spaces left, and the beta closes on 4 July.

Most schools have never actually checked what their AI use is teaching students to do, so they cannot see the gap, let alone close it. That is where the new addition to the audit helps: a full resource of twenty-four packs, editable policy templates, CPD decks, safeguarding tools and inspection evidence, mapped to a Leading AI timeline, so no leader has to start from a blank page.

The part I did not plan: I edited most of those packs using Claude Fable 5, the top-tier model Anthropic put into general release on Tuesday. By the weekend it was gone, pulled worldwide three days after launch under a US export-control order that barred access by any foreign national. Sitting in Dubai, I am precisely the foreign national that order names. I finished the build on Friday. One day later I could not have opened the model at all. Anthropic calls it a misunderstanding and expects to restore access. Possibly. I would not build a deadline around possibly.

I keep coming back to a line from Adam Grant. The hallmark of expertise is no longer how much you know. It is how well you synthesise.

I spoke at the Google for Education Leaders Conference and boiled leading with AI down to two moves: name the mechanical and keep the thinking. Naming the mechanical means saying out loud which work the tool should take, the drafts, the notes, the scaffolds, the summaries. Keeping the thinking means protecting the moments where judgement has to stay human, and building friction back in there on purpose.

I am still unpacking how the OUP report sits next to MIT’s findings. Students using AI without scaffolding develop cognitive debt. Students given access to AI in a carefully designed school task decline it seventy-two per cent of the time. Both are true. The variable is the design of the environment.

For the longer-form companion to the OUP report, read Parker et al., Longitudinal insights into AI in education, published this year in Computers and Education Open. It tracks more than three hundred trainee teachers across four semesters in a US programme. Use of AI for assessment preparation rose from fifty-seven to eighty-three per cent; use for studying from forty-four to seventy-six per cent.

The more interesting finding sits underneath: a statistically significant difference in ethical reasoning between AI users and non-users in the same cohort. These are the people about to enter classrooms as teachers. Worth fifteen minutes.