A school AI tutor is a piece of software used by children. That means safety, privacy, accessibility and the curriculum are engineering requirements, not a paragraph someone adds at the end.
What we mean by an AI tutor is set out on the AI tutor for schools page.
Education is a software problem
I am guest-editing a special issue of IEEE Software with Xiwei (Sherry) Xu and Mahsa McCauley, on engineering responsible AI for education. The call for papers is still to come. The point of the issue is already clear.
AI in education is often talked about as a conversation. It can also sit quietly behind existing conversations between learners, teachers and the resources they already use. Designing that well means engineering for safety, privacy, accessibility and curriculum grounding. It means putting children's rights at the centre, leaving room for teachers and parents to see what is happening, and evaluating whether anyone actually learned more.
The issue is meant to cover schools, tutoring, vocational education and universities. The shared question is a practical one. What does it take to build these systems so they can be trusted?
A curriculum a machine can read
John Roberts, chief executive of Oak National Academy, talked on the AI in Education podcast about the public, machine-readable model Oak has built of the English national curriculum.
A curriculum published as documents is written for people. Software cannot use it reliably, so each company builds its own model, and those models do not always agree. Oak's ontology is open infrastructure that anyone can build on. We would not have launched in England as soon as we did without it. I wish every system did this. A school AI tutor that has to guess what the curriculum says will guess differently from the one next door.
A coding agent can save the secret as well as the command
Apiiro found database passwords, API keys and cloud credentials sitting in public GitHub repositories. In Claude Code, choosing "Yes, and don't ask again" for some commands can save the whole command, credentials included, in .claude/settings.local.json. Commit that file and the secret goes with it.
Anthropic has acknowledged the issue and is looking at further safeguards. Anyone using the tool should check their repositories and history, ignore that file, and rotate anything that leaked.
Schools are starting to let staff and older students use coding agents. The lesson is the same one as for a school AI tutor with tools. The dangerous moment is not the clever answer. It is the permission that outlives the conversation.
Who is paying whom to buy the chips
Colette Kress is interim chief executive at Nvidia, and the interesting question is what happens to the company's investment strategy. Under Jensen Huang, Nvidia became more than a chip supplier. It invested in customers and backed the construction of data centres that then buy Nvidia chips.
There are ordinary commercial reasons to do that. It also creates a loop: the supplier helps finance the demand for its own product. If the demand those investments assume does not arrive, the risks are shared across the companies in the loop. Reuters has reported investor concern that Nvidia could become less willing to keep playing that role.
A school deciding whether to adopt a school AI tutor should notice the difference between a tool that is being paid for because it works, and an industry being paid for by the people selling the infrastructure.

