A department paid to train AI tools on teachers' own lesson plans

The classroom note

A teacher marking a stack of essays on a Sunday evening is the audience the Department for Education named directly in its 28 August 2024 announcement of a 'content store': a pooled bank of curriculum guidance, lesson plans and anonymised pupil assessments, built with the Department for Science, Innovation and Technology and the Open University, and made available to education technology companies to train their AI tools on. The stated goal, in minister Stephen Morgan's words, is to 'reduce out of hours work and boost time spent teaching children,' with £3 million funding the store itself and £1 million offered as prizes to companies building working tools by March 2025.

What the evidence says

The August announcement makes a claim about mechanism, not measured outcome: training an AI system on curriculum-specific data should, in principle, produce more accurate results than training on generic web text. A follow-up announcement from 13 January 2025 puts a number on that principle, stating feedback accuracy reached '92%, up from 67% when no targeted data was provided.' That figure is not sourced to a named independent study; it is presented as the government's own comparison, without sample size, subject area or who scored 'accuracy.' The same release states developers, not an evaluator, 'estimate' some tools 'could save time spent on formative assessment by up to 50%,' a projection attributed to the tool-builders rather than measured classroom use.

The implementation question

The mechanism worth naming plainly is licensing and access: the content store pools public curriculum material and anonymised pupil work so private companies can train commercial products on it, in exchange for £1 million in prize funding for sixteen edtech developers building prototypes due by April 2025. That is a data-access decision with implications for who benefits from public-sector content, not only a workload fix. The training cost for a school is close to zero; the exposure is that pupil work, even anonymised, becomes training data for products a school does not select and cannot audit beforehand.

What holds and what fails

What holds is the underlying logic that grounding a tool in curriculum-specific material should reduce the kind of generic, ungrounded error generative AI is known for; the 92-versus-67 figure at least gestures at a mechanism for improvement, even without independent verification. What fails, or has not yet been shown, is the 50 percent time-saving claim, which the department's own release attributes to developer estimates rather than an evaluated pilot, and readers should treat it as a vendor projection until an independent study of the prototypes due by April 2025 is published. This is an editorial caution: a funding announcement is not evidence a tool works.

• Which of the sixteen funded tools have published an independent evaluation since April 2025?
• What anonymisation standard was applied to the pupil assessments in the content store?
• Does the school's own data-protection policy address content it did not know was used for AI training?

A content store changes what AI tools can be trained on; it does not, by itself, change what a tool actually gets right in a classroom.