AI in schools: the evidence test for buy, trial or stop

01 / Controlled deploy, trial or stop

“Buy” is the end of a decision process, not a category of technology.

DfE says teacher-facing generative AI offers more immediate benefits and fewer risks than pupil-facing use, while the professional and organisation remain responsible. It also says evidence remains limited on learner development, educational outcomes and child safety.1 DfE’s 13-area safety standard widens the test beyond privacy and filtering to cognition, relationships, mental health and manipulation.2

The standards are a non-statutory benchmark, not certification, an approved-product register or proof of learning impact.2 Ofsted neither requires AI nor evaluates products as a stand-alone inspection item; it says the evidence is still insufficient to define “good” use.10

02 / Why 2026 is different

An old AI checklist no longer covers the current decision.

• Product safety widened: DfE’s January 2026 update expanded its mainly supplier-facing product safety guidance to 13 areas, including cognitive development, emotional and social development, mental health and manipulation. DfE says schools and colleges may use the standards as a product-safety benchmark; they are not certification or proof of impact.2
• Data law changed: all data-protection provisions of the Data (Use and Access) Act 2025 were in force by 19 June 2026. It amends rather than replaces UK GDPR and the Data Protection Act. DUAA opened the full range of lawful bases for significant automated decisions where appropriate safeguards continue to apply; special-category data remains subject to stricter conditions.7
• Provider evidence matters: ICO reported 28 consensual edtech audits and 596 recommendations, with recurring gaps in roles, contracts, data maps, minimisation, retention, privacy and DPIAs. This is not a prevalence estimate.8
• Assessment boundaries sharpened: Ofqual expects awarding organisations to demonstrate valid, reliable and fair outcomes, with appropriate expert human involvement, quality assurance and accountability; it says AI may not be the sole marker in any assessment that forms part of a regulated qualification.11 Its January 2026 working paper adds that validity evidence must fit the assessment context and that agreement with human marks alone is insufficient.22
• Agentic systems reached the risk register: NCSC calls for bounded pilots, least privilege, temporary credentials, logging, sandboxing, active oversight and a stop mechanism. Its August 2026 advice is interim.199
• School standards became more AI-specific: DfE updated filtering, cyber, governance and accessibility standards on 25 August 2026.5161718
• Leadership support was refreshed: DfE updated its school and college leader materials for 2026/27 on 19 May. They support audit, strategy and implementation; they are not effectiveness evidence or product approval.24

03 / The five-gate decision framework

Any red gate means stop. Amber means remediation or a bounded trial, not silent acceptance.

The five gates are School Connection’s synthesis of the cited law, statutory guidance, regulator positions, non-statutory standards and official evidence. They are not a DfE or regulator-issued approval scheme.

DfE’s current school data and EdTech guidance supports this whole-chain view: lawful basis, minimisation, roles, sub-processors, contracts, security, retention, deletion and exit all sit inside the decision.34 For academy trusts, a framework can streamline procurement, but it does not remove the trust’s responsibilities as contracting authority. Neither a framework nor an approved buying option establishes AI safety, data-law compliance, accessibility, educational effectiveness or value for the intended use.25

Equality is not a supplier checkbox. Relevant public authorities in Great Britain must exercise due regard under the Public Sector Equality Duty before and during the decision and continue to monitor effects; the duty cannot be delegated to a vendor or an impact-assessment document.21 DfE’s accessibility standard also places accessibility evidence, assistive-technology compatibility and accessible procurement inside the decision.18

04 / The pupil-facing test

Child safety is more than filtering harmful words.

Under this article’s decision framework, pupil-facing tutors, chatbots, adaptive practice and writing aids should default to supervised, time-limited trials in 2026. DfE requires close supervision and appropriate age, safety, filtering and monitoring controls; the “trial” classification is School Connection’s evidence-led recommendation, not a statutory rule. The governance conclusion is strong; general product-effectiveness evidence is not.

• Age-appropriate filtering, usable logs and a named local safeguarding alert route
• Close supervision, age compliance and controls that work for SEND and EAL contexts
• Progressive disclosure and prompts for the pupil to attempt work before receiving an answer
• Measures of cognitive offloading rather than completion alone
• No implied personhood, emotional dependence, sycophancy or manipulation
• Human routes for distress, session limits and no commercial steering

These expectations come directly from DfE’s education policy, product standards and filtering-and-monitoring standard.125 They are reasons to test carefully, not evidence that a product meeting them improves learning.

05 / The data map

If the organisation cannot draw the chain, it cannot approve the use.

A DPIA must be completed before processing that is likely to result in a high risk to individuals’ rights and freedoms; AI alone does not automatically trigger one. Screen every use and document the reasons where a DPIA is not required. Children, profiling, systematic monitoring, sensitive data and innovative technology can combine to trigger the requirement.6 For a decision not to count as solely automated, human involvement must be meaningful and active rather than token. Where a solely automated significant decision is permitted, safeguards include information, a way to express a view, the ability to obtain human intervention and a challenge route.20

Contract labels are not conclusive. ICO says the facts and purposes determine whether an edtech provider is a processor, controller or joint controller and whether the Children’s Code applies to its service.26

06 / Human oversight

“A human is in the loop” is not a control unless the human can change the result.

For a consequential output, the reviewer must be competent, see the relevant evidence, have enough time, understand the system’s limits, be able to reject or change the result and record a reason. Where UK GDPR articles 22A to 22D apply, safeguards include information, a way to express a view, human intervention and a challenge route.720

The same principle holds in assessment. Ofqual’s position requires expert human involvement and prevents AI acting as the sole marker in a regulated qualification assessment.1122 A low-stakes classroom feedback trial does not establish fitness for a high-stakes qualification.

The sole-marker boundary applies to assessments that form part of Ofqual-regulated qualifications; it is not a blanket prohibition on every low-stakes classroom feedback tool. AI-generated work submitted as a pupil’s own can be malpractice, so leaders should map coursework and non-exam-assessment exposure, set task-specific rules and follow current awarding-organisation and centre procedures when authenticity is in doubt.23

07 / A trial that can fail

Pre-register the decision before the excitement of the pilot changes the question.

Fix the users, site, subject, task, version and duration. Set a baseline and comparator. Choose one primary benefit and non-inferiority thresholds for quality, safety and equality. Measure net workload after checking and support; use blind quality review; test unaided learning; examine relevant subgroups; log incidents and near misses; count direct and hidden cost; and name the reviewer, decision date and stop rules before launch.

Questions for leaders and trustees

09 / What we are monitoring next

AI approval is not permanent.

Reopen the gate record after a material model, feature, service-plan, supplier-term, integration, data-category, user-group or autonomous-action change; after an incident, repeatable jailbreak, accessibility complaint or subgroup disadvantage; when law, KCSIE, regulator or awarding-organisation rules change; when NCSC replaces its interim agentic guidance; and at contract renewal or a claimed impact milestone.