A government record disclosed how a public lesson AI works

The classroom note

A teacher asked to try a new AI lesson-planning tool is usually handed a product page, not an engineering account of how it works. Oak National Academy, a curriculum body funded by the Department for Education, published a government algorithmic transparency record for Aila, its AI lesson assistant, dated 17 December 2024. It states the tool was in 'Public Beta' at that point, runs on OpenAI's GPT-4o alongside a reranking model, and works by retrieval-augmented generation: a search across a vectorised corpus of roughly 10,000 of Oak's own lessons, so a generated plan is built from resources designed and tested by teachers and subject experts, as the tool's own product page puts it, not from an open-ended model response.

What the evidence says

The transparency record is unusually specific for this product category: it names the exact model version, states the corpus contains 'no personal data, protected characteristics or proxy variables,' and describes evaluation using large language models as judges against Oak's own pedagogical rubric, cross-checked with teacher feedback. That is a disclosed methodology, not an independent trial; the auto-evaluation is Oak assessing its own tool by its own rubric, and the record reports no external accuracy figure and no comparison against unassisted lesson planning. Oak's homepage adds only that the tool is free and that Oak is the first to admit it's not perfect, an admission alongside, not a replacement for, the record's technical detail.

The implementation question

The record states humans remain in the loop at every stage, that an independent, asynchronous moderation agent scans outputs for unsafe content, and that the tool is deliberately not designed to generate a finished lesson 'at the touch of a button.' That is the risk mechanism made explicit: grounding in a fixed curriculum corpus is meant to reduce fabrication, and moderation plus mandatory teacher editing is meant to catch what grounding misses. The cost side differs: the tool is free to use, but the content store and DfE funding behind Oak are public spending, and a school's real cost is teacher time spent reviewing and editing every output, not a licence fee.

What holds and what fails

What holds is that grounding output in a fixed, vetted corpus narrows the space for invented facts more than an open general-purpose assistant would. What is not established is whether teacher editing in practice catches every error the moderation agent misses, or whether accuracy gains claimed for other content-store-trained tools apply to Aila specifically; the record makes no such claim for Aila. Editorially, a beta label and a published transparency record are a stronger basis for trust than marketing alone, but not a substitute for a published outcome study.

• Has the beta label been removed, and if so, what changed in the underlying model or evaluation?
• Does the moderation agent's scope cover subject-specific factual errors, or only unsafe content?
• What proportion of a generated lesson does a teacher typically need to edit before using it?

A transparency record is a disclosure, not a guarantee; it tells a school what to check, not that checking is unnecessary.