It’s Sunday evening. A teacher sits down to plan a Year 8 science lesson and tries an AI tool to help save some time. The output looks impressive. It’s fluent, well-structured and authoritative.

But something isn’t quite right. It doesn’t build on what her class covered last term. It introduces concepts they haven’t encountered yet, and it bears little resemblance to what the national curriculum actually requires her to teach. She spends the next 40 minutes fixing it. So much for saving time.

This is not an isolated experience. Three in four teachers are now using AI for their day-to-day work, according to the NEU, and many are having exactly this experience.

The problem here isn’t the tool. It’s that the AI has a limited grasp of what the national curriculum actually says, and even less idea how it all connects. This is because the national curriculum, in its current form, is almost impossible for a machine to use.

England’s national curriculum lives in PDFs and prose documents - useful for humans, but deeply limited as a foundation for AI.

Large language models can extract some meaning from text, but not enough to know what a Year 8 science class should already understand, what they’re working towards or how one concept connects to the next.

The first problem is about access to digital curriculum content. The second, and harder, problem is an understanding of how that content fits together.

Most AI tools in schools have neither. The ones teachers can trust need both. That gap between what AI could do for teachers and what it currently does is a problem that Oak National Academy has been quietly working to fix.

Mapping the curriculum

So what have we been doing? In simple terms, we’ve been building knowledge graphs - maps of the entire curriculum that capture not just what needs to be taught but how everything connects.

Think of a maths teacher introducing Pythagoras’ theorem. If a pupil is struggling, the teacher knows it’s because they haven’t yet secured their understanding of angles or triangle properties. He adjusts his teaching accordingly.

That kind of professional knowledge is what experienced teachers carry around in their head. Knowledge graphs take those connections, dependencies and sequences and translate them into a form that AI tools can actually read and use. It’s precisely what most are currently missing.

Knowledge graphs are not AI and not curriculum. They don’t tell teachers how to teach. They won’t guarantee a good lesson. Most teachers will never see them.

But combine them with rigorous safety, accuracy and bias checks, and you start to get AI that teachers can genuinely rely on, grounded in what pupils actually need to know, and when.

What Oak has built, and why it’s free

Oak has now mapped six subjects as curriculum knowledge graphs: maths, English, science, history, geography and citizenship. More are in development.

This work isn’t just about trying to make Oak’s own AI tools better - the knowledge graphs are freely available to everyone, and we’ll be making sure they can be plugged into the main AI tools teachers are using: ChatGPT, Claude, Copilot and Gemini.

It’s about raising the floor on quality for every tool in the market, so curriculum-aligned AI becomes the baseline expectation, not a premium feature.

There is another important benefit. Most schools will end up using more than one AI tool for different activities - perhaps one for planning and one for assessment. When different tools are built on the same foundation, they fit together, so teachers don’t end up with different tools telling them different things.

That’s one of the fundamental building blocks of a sector where AI is trustworthy.

The questions to ask

Most schools are navigating AI adoption without much infrastructure to support them. Knowledge graphs won’t replace the need for clear policies and guidance, but they raise the floor on quality for every tool that uses them.

So when an AI tool arrives in school, there are two questions worth asking. Firstly, does it know what the national curriculum says? And secondly, does it understand how the curriculum connects and sequences, or is it working from a general model trained on the internet and hoping for the best?

The first question is a minimum requirement; the second is what makes the difference between an AI that generates a plausible-sounding lesson and one that generates the right lesson for the right class.

Infrastructure is rarely an exciting story. No one turns the page to read more about data schemas.

But get the infrastructure wrong and it doesn’t matter how impressive the AI looks; it risks confidently leading teachers in the wrong direction. Get it right and every tool built on top of it becomes more trustworthy, more useful and more worth a teacher’s valuable time.

John Roberts is interim CEO of Oak National Academy