School leaders are asked to make decisions about AI tools that change every few months, using research that takes years. On 24 September the Department for Education published two documents that take that problem seriously.
The first is a report from the DfE Science Advisory Council's AI and digital education working group, setting out eight principles for evidence-informed policymaking in fast-changing technology contexts. It is advice to the department, not DfE policy, but it reads like a practical checklist for anyone choosing an AI tool. The second is guidance for a new EdTech Testbed Programme: £24 million over four years to test education technology, including generative AI tools, in more than 1,000 schools and colleges in England.
The evidence problem, stated plainly
The report cites a recent Stanford meta-analysis of approximately 800 papers on AI in education that found only 20 studies meeting the threshold for causal evidence. That is the evidence base schools are working with.
The report's response is not to wait. It argues that demanding randomised controlled trial standards before giving teachers any guidance “is not a neutral methodological choice”. In a fast-moving technology context, it says, that demand “functions as a decision to withhold guidance indefinitely.”
Instead, it proposes treating evidence as something that builds up over time from several sources: research, usage data, practitioner experience and the lived experience of pupils and teachers, each treated as partial.
Four principles a school can use now
Several of the eight principles translate directly into school decisions.
Keep options open. The report says decisions should avoid irreversible commitments before there is enough evidence about long-term effects. It warns specifically about procurement that creates monopolies, proprietary lock-in or dependence on a single provider. For a school or trust, that means asking how easily you can leave a contract, and whether your data and materials come with you.
Evaluate conditions alongside products. A tool may work in one school and not another, depending on teacher training, the curriculum and how closely adults supervise its use. Products change within months; the conditions change more slowly. Before asking whether a tool works, ask whether your school has the conditions in which it could.
Build in harm detection from the start. The report names cognitive offloading and erosion of independent judgement as harms that need evaluating from the outset. It notes that short-term performance gains during use can mask longer-term harm to learning that only shows when the tool is removed. That is a question to ask of any pupil-facing AI tool: what happens when pupils work without it?
Ask for usage data. The report suggests requiring developers to provide a consistent minimum set of usage data, in an accessible format, as a condition of public sector procurement. Schools cannot impose that alone, but they can ask now. If a supplier cannot tell you how often pupils use the product and for what, you cannot judge whether it is helping.
What the testbed will add
The EdTech Testbed Programme will run in two stages. First, a consortium led by Ipsos will carry out up to 100 rapid evaluations of products and practices. The most promising will move to a second stage, where the Education Endowment Foundation will invest £7 million in up to 15 randomised controlled trials or other rigorous evaluations.
Schools, colleges and suppliers will be recruited in phases from the 2026 to 2027 academic year. Results will be published through an online evidence hub. That will take time, and it covers products more than the AI literacy pupils need. But it will give schools in England something they rarely have: independent evidence about specific tools, gathered in settings like theirs.
A better set of questions
Taken together, the two documents suggest a more useful conversation with any AI supplier. Not just “Does it work?” but: In what conditions has it worked? What usage data will you share with us? How would we know if it was doing pupils' thinking for them? What happens if we stop using it?
A school does not need a perfect evidence base to make a sound decision. It needs to keep learning after the decision is made.
Sources: DfE, Evidence-informed policymaking and fast-changing technology, 24 September 2026; DfE, EdTech Testbed Programme: guidance, 24 September 2026. The principles are independent advice from the DfE Science Advisory Council working group, not government policy. Both documents apply to England.