Curriculum is the missing data layer of education AI

Almost every demonstration of AI in education begins with the same implicit assumption: that the curriculum the system is supposed to teach already exists in a form a machine can reason about. A student opens a tutor. The tutor knows where the student is in the curriculum, what they have covered, what comes next, how this concept builds on previous ones, what assessment evidence would indicate mastery, and which adjacent topics to bring forward if the student is struggling.

None of that is true in any system in production today. The curriculum the tutor is supposedly teaching against does not exist as data. It exists as a PDF on a ministry website, a framework document in a teacher's drawer, a set of textbook chapters with non-aligned numbering, and — most of all — as tacit knowledge in the heads of the teachers who have taught it for years.

Why this is the central problem, not a side problem

The AI conversation in education has spent three years discussing models. Whether the model is large or small, whether it is hosted or local, whether it is fine-tuned or prompted, whether it hallucinates, whether it can reason. These are real questions. They are also downstream questions.

Upstream of every model is the data the model is supposed to operate against. In every other domain where AI has produced durable value — clinical decision support, legal research, financial analysis, industrial diagnostics — that domain spent a decade or more building a structured data layer before AI became useful on top of it. Hospitals built coded clinical terminologies. Law firms built citation graphs. Banks built normalised transaction models. Manufacturers built telemetry schemas.

Education has not done this. The national curricula of most countries — including most of the world's largest education systems — are not machine-readable in any operational sense. They are documents. They have structure for humans to follow and almost no structure for machines to reason against.

What "structured curriculum" actually requires

A curriculum that an AI system can reason against is not a digitised PDF. It is a graph of educational concepts with explicit relationships:

• Each learning outcome expressed as a discrete, addressable unit
• Prerequisite relationships between outcomes — what must be known before what
• Progression relationships — how outcomes build into larger competencies
• Assessment evidence specifications — what would constitute mastery of each outcome
• Cross-references to recommended pedagogical approaches and exemplar tasks
• Mappings to textbooks, content libraries and assessment items already in use
• Versioning and provenance — which authority specified this, when, and under what framework

Building this for a single subject in a single country is a multi-year programme of work involving curriculum specialists, classroom teachers, assessment experts and engineers. Building it across subjects, year groups and national systems is the kind of infrastructure project that quietly underwrites entire categories of technology for decades.

Why frontier labs will not build this

Structured curriculum intelligence is exactly the kind of asset frontier AI labs have no incentive to construct. It is slow. It is jurisdiction-specific. It requires deep, sustained domain partnership. It scales linearly with effort, not exponentially with compute. It does not improve benchmark scores. And the customers who care most about it — ministries, regional education authorities, large public school systems — are not the customers frontier vendors are commercially organised around.

This is not a temporary gap that a larger model will close. A model with a trillion parameters and a perfect reasoning score still cannot answer the question "what does the Year 7 algebra curriculum in this country actually specify, and how does it progress into Year 8?" without that information existing somewhere as structured data. The model is the wrong layer of the stack to ask.

What becomes possible once it exists

The moment curriculum exists as structured intelligence, a long list of things that have been promised by education-AI products for a decade actually become deliverable.

Lesson planning stops being content generation and becomes curriculum-aligned sequencing. Assessment stops being question generation and becomes targeted evidence collection against specified outcomes. Personalisation stops being adaptive difficulty and becomes progression-aware pathway construction. Reporting stops being engagement metrics and becomes mastery against defined competencies. Cross-system insight — what is being taught, where, how well, and where the gaps are — becomes a query, not a research project.

None of these require a more powerful model than already exists. They require the data layer beneath the model to exist at all.

Why this is a defensibility question, not just a feature

Structured curriculum intelligence has the same shape as every durable data moat in technology history. It compounds with use — every deployment refines the encoding. It is jurisdictionally specific — what works for England's Key Stage 3 mathematics does not transfer to Indonesia's Kurikulum Merdeka. It requires direct relationships with the bodies that actually own the curriculum — ministries, examination boards, framework authorities. And it grows linearly with effort, which means whoever starts earliest accumulates an advantage that cannot be closed by capital alone.

The companies that build this layer first, for each major education system, will be the companies the next decade of educational AI is built on top of — including by the frontier labs, which will eventually need to call into it.

"Every demonstration of AI in education today assumes the curriculum exists as data. It does not. The organisations that build the structured curriculum layer — system by system, country by country — are building the substrate the entire category is missing."

aime is building that substrate.