Accessible by default, not accessible on request.

An accessible learning resource is one that works without anyone having to ask for a different version. In practice that means readable type, real structure, described images, plain sentence construction, and a format that a screen reader or a symbol board can actually parse. AI can produce all of that quickly, which changes the economics of accessibility far more than it changes the principles.

What makes a learning resource accessible?

Structure first, appearance second. A document with real headings can be navigated; a document where headings are just bigger bold text cannot, however identical it looks on paper. The same is true of tables used for layout, images carrying meaning with no description, and colour used as the only way to distinguish one thing from another.

• Real heading structure, in order, so the document can be navigated rather than only read.
• Descriptive alternative text on every image that carries meaning, and none on decoration.
• Contrast that holds at the level WCAG requires, checked with numbers rather than judged by eye.
• Sentence construction the reader can decode: one idea per sentence, active where possible, no idiom doing load-bearing work.
• A layout that survives being enlarged, reflowed or read aloud, because plenty of pupils will do at least one of those.

How can AI help make materials accessible?

By making the second version free. Historically the reason resources were not accessible was not ignorance, it was cost: producing a symbol-supported version, a large-print version and an easy-read version of the same worksheet meant three more jobs on a Sunday. When the marginal cost of the fourth version drops close to zero, the argument for producing only the standard one disappears.

It is also good at the mechanical parts people skip. Drafting alt text for thirty images. Rewriting a passage at a lower reading age without dropping the content. Producing a plain-language summary of a policy for a parent who will not get through four pages of it. None of that is intellectually hard. All of it is exactly the sort of work that never gets done.

Where does AI get accessibility wrong?

Confidently, and in ways that look finished. Generated alt text often describes what an image contains rather than what it means, which is useless in a diagram where the meaning is the relationship rather than the objects. Simplified text frequently loses the qualifier that made the sentence true. Symbol support generated without knowledge of the pupil's actual symbol set is not support, it is decoration.

There is also a quieter failure. A tool that offers an accessible version as an option has reproduced the original problem in software: someone still has to ask, and the pupil still has to be the reason. Accessibility as a checkbox is a different product from accessibility as a default, and they look similar in a demonstration.

Why we build this way first

Aevum is built accessibility-first rather than accessibility-eventually, because the founder spent two decades watching who gets left out when technology is designed for the average child, and has been on the receiving end of that design elsewhere. The practical consequence is that the accessible output is the output, not a variant reachable through a settings panel.

It also means the accessibility work is testable and published rather than asserted. Contrast is checked with real numbers. Screen reader behaviour is a requirement, not a hope. In development, and the Trust Hub carries what we have committed to so far.