On 24 September 2026, the UK Department for Education published an eighteen-page report produced by the 'AI and digital education' working group of its Science Advisory Council. The document, 'Principles for evidence-informed policymaking in fast-changing technology contexts', starts from a timing problem: a scientific publication dated 2023 or 2024 is generally based on data collected in 2021 or 2022, when uses, capabilities and products have already changed.

'The demand for randomised controlled trial level evidence before any recommendation to practitioners is not a neutral methodological choice', the working group writes. 'In a fast-moving technological context, it functions as a decision not to give recommendations, indefinitely.' The text is independent advice submitted to the department, it does not commit the department's policy, and the GOV.UK page specifies that it applies to England.

Eight principles grouped into three families

The first family concerns the logic of evidence. The report calls for combining and triangulating sources, treating each as partial, adjusting evidence requirements and standards to the question asked since improving a product, advising a teacher and detecting harm do not call for the same methods or the same timescales, and treating collection as a continuous and formative process, on the model of formative assessment of a pupil.

The second family concerns methods: keep options open, evaluate conditions of use at the same time as products, adapt the method to the maturity of the tool. The third concerns infrastructure: build systems that allow evidence to accumulate, and make it accessible to the audiences for whom it is intended. The whole is placed under a label, 'Bayesian accumulation of evidence', which the five contributors are careful to define: it 'does not mean lower requirements', it requires 'combining inputs honestly, revising openly and making assumptions explicit'.

Two figures support the demonstration, both cited in the report. A Stanford meta-analysis of around 800 articles devoted to AI in education retained only twenty studies crossing the threshold of causal evidence. A review of UK research projects funded on the subject found no project using randomised trials as the main evaluation method. The council maintains this method where tools are stable, where the learning mechanism is demonstrated or updates are minor.

Usage data required as a condition of public procurement

The most operational part of the text concerns public procurement. Principle 4 calls for reversibility: 'Procurement structures that create monopolies, proprietary lock-in or dependence on a single supplier before evidence is available must be treated with particular caution.' The cost of a change of course, the contributors recall, remains low as long as a technology is not installed, and rises sharply once it is.

Principle 7 shifts the burden of demonstration to the supplier. It proposes, among possible routes, requiring developers to provide 'a coherent set of usage indicators in an accessible format, with an agreed minimum data base, as a condition of public procurement'. It would then be up to the supplier to establish, through this continuous provision, that its product is safe and effective in use. The report adds a second layer: a structured platform where teachers would share their experience of tools in the classroom, 'closer to a peer review mechanism than a consumer review site'.

Six months earlier, the CSEN said how to prove before deploying

France has a comparable body. The Scientific Council for National Education (CSEN), chaired by Stanislas Dehaene, held an international conference on 25 March 2026, 'What uses of AI in education?', opened by Minister Édouard Geffray, and published its conclusions and recommendations in April.

Its first recommendation to the institution can be summed up in one word: wait. The companies developing these tools have 'often an insistent dissemination policy towards teachers and pupils', the council observes, and 'faced with this pressure, it is essential to be patient and to experiment'. Four requests follow: assess whether pupils' progress is real, 'without simply collecting their impression, that of teachers, even less that of the company'; require companies to 'go through several validation stages, up to full-scale testing'; make explicit whether the use of AI is authorised; finally reserve this use for situations where these tests have demonstrated a 'proven pedagogical added value'.

The CSEN puts this last phrase in quotation marks. It already appears in the framework for the use of AI in education published by the ministry on 13 June 2025, whose summary page enjoins: 'Use AI only when pedagogical added value is proven.' The scientific council's contribution concerns the method, and it designates the measuring instrument: 'The national assessment system is fully adapted for this.'

Two evidence timelines, one text that conditions use

The two councils want the same thing, to have the demonstration borne by the supplier. They diverge on when to decide. On the French side, prior demonstration conditions use, and it comes from the ministry, not the scientific council. On the British side, this same condition applied to products that change 'within months' amounts to never advising.

Neither opinion has changed the applicable law. The official position of the UK Department for Education remains the page 'Generative artificial intelligence (AI) in education', updated on 12 August 2025, which announces the pilot of an 'edtech evidence board' responsible for assessing the quality of impact evidence for educational tools. The French framework, for its part, was published after a national consultation conducted from January to May 2025. Its twenty-two pages also stand out by age: awareness-raising from primary school without manipulation of generative AI, pedagogical use 'authorised in class from year 9 (4e)', compulsory training on Pix at least in year 9, year 11 (seconde) and first year of CAP. ActuIA had detailed its content in June 2025, then analysed the Calvez report, which proposed on 1 July 2026 to lower this threshold to year 7 (6e).

No page of this framework sets any public procurement rule. Its only tool selection instructions amount to two lines: favour open-source solutions, and use AI only if no other less ecologically costly solution meets the need.

For a secondary school, a region, a publisher

The shift proposed by the British council carries a barely visible cost. Giving up the randomised trial as the sole standard assumes that someone collects continuously, and the report places the exploitation of this data and the sharing of experience on schools and teachers. Principle 7 admits that the necessary infrastructure 'does not yet fully exist', and the text nowhere indicates who funds this work or how a secondary school learns to read usage traces. The CSEN, for its part, bases measurement on a system that already exists, national assessments, and leaves the school without an answer on the day a publisher offers it a tool that no full-scale test has yet passed.

For a local authority that purchases, region or department, the English text offers a clause that can be transposed as early as the next tender: condition the award on the provision of usage indicators and the maintenance of reversibility. For a French publisher, it announces the question that public buyers will ask, which will concern as much the data it agrees to make consultable as the demonstrated effectiveness of its product.

As of 28 September, four days after its publication, the GOV.UK page of the report shows no update and the department has attached no response. The same morning in Paris, the historical CSEN website, www.csen.education.gouv.fr, to which Réseau Canopé redirects its links to the recommendations, was inaccessible in a browser: its security certificate expired on 27 September at 11:27 UTC.

Our articles will then appear first in Google Top Stories.