Education AI: Keep Learning Teacher-Led

An AI tutor can produce another explanation. It cannot know whether a pupil is quietly confused, whether the curriculum sequence is appropriate, or whether an apparently correct answer reflects understanding. A generated worksheet can save drafting time and still contain a subtle misconception.

The safe operating principle is simple: AI may support the learning relationship; it must not become the relationship or the final educational judgment. Teachers define the purpose, select material, observe learning and decide what happens next.

This guide reflects authoritative sources available on 31 July 2026. Education, safeguarding and qualification systems are devolved. Department for Education, Ofqual and Ofsted sources cited below generally apply to England. Schools and colleges in Scotland, Wales and Northern Ireland must use their own government, regulator, awarding-body and safeguarding guidance. Higher education also has different governance and assessment arrangements from compulsory education.

Define the Learning [Contract](/industries/legal)

“Personalise learning” is too broad to evaluate. State the learner group, subject, task, teacher role, information used, permitted output and evidence of success.

Every interface should tell pupils when they are interacting with AI, what it can and cannot do, how to reach a person and whether the conversation is retained. A tool must not imply that it is a teacher, counsellor or safeguarding professional.

Start With Teacher-Facing, Low-Consequence Work

The first pilot should remove a bounded drafting task, not mediate a vulnerable child’s learning or make a grade decision. Suitable candidates include creating a first draft of retrieval questions from approved curriculum material or reformatting a teacher-written explanation.

Use a source-grounded workflow:

• teacher selects the approved curriculum source;
• tool generates a draft with citations or source references where supported;
• teacher checks subject accuracy, difficulty, language, inclusion and answer key;
• edited resource is versioned separately from the model output;
• pupils receive only the approved version; and
• the team records preparation time and correction types.

DfE’s generative AI policy for education says education technology may offer opportunities but staff must protect personal data and intellectual property and maintain professional judgment. Its 2026 leadership materials provide an audit and planning route for schools and colleges in England.

Do not assume time saved. Measure total preparation, checking, correction and rework. A faster first draft can create slower quality assurance. Compare against the existing process over several weeks and subject types.

Make Tutoring Reveal Thinking, Not Supply Answers

An adaptive tutor should diagnose the *next useful question*, not race to finish the task. Design interaction around curriculum-aligned steps:

• ask the pupil to attempt before revealing help;
• request an explanation, working or evidence;
• offer the smallest useful hint;
• use teacher-approved examples and vocabulary;
• surface persistent misconceptions to the teacher;
• stop repetition that is causing frustration;
• preserve an easy “ask my teacher” route; and
• avoid emotional dependency, praise that manipulates continued use or claims of friendship.

Do not infer ability from response speed, spelling, accent or device behaviour without evidence and a legitimate educational purpose. A pupil may be using assistive technology, sharing a device, learning in another language or working with interrupted connectivity.

DfE’s January 2026 generative-AI product safety standards cover educational purpose, filtering, privacy, security, intellectual property, testing, cognitive development, emotional and social development, mental health and manipulation. Use them as procurement acceptance criteria, then test the installed configuration with the school’s pupils and curriculum.

Government announced teacher-supervised development of new tutoring tools in April 2026, with tools intended for 2027 after co-design and testing. That timetable is a useful caution: a funding announcement is not evidence that current generic chatbots deliver safe, effective individual tutoring.

For early-years settings, use the age-appropriate controls in our AI childcare guide.

Demand Evidence of Learning, Not Engagement

Minutes used, messages sent and questions completed can rise while learning remains unchanged. Establish an educational evaluation before rollout.

Define:

• primary learning outcome and when it will be measured;
• baseline and comparison group or credible counterfactual;
• pupil groups and minimum sample needed for useful interpretation;
• implementation measures, including actual usage and teacher training;
• adverse outcomes such as misconceptions, dependence or exclusion;
• data-analysis plan agreed before results are seen; and
• who can independently review the evidence.

Use curriculum-valid measures outside the tutor itself. If the product teaches and tests with the same item style, score gains may reflect familiarity with the interface rather than transferable understanding.

An Education Endowment Foundation 2026 intelligent-tutoring trial illustrates the right level of specificity: a named Key Stage 2 maths programme, implementation requirements, independent evaluation and a defined period. It does not prove that every adaptive or generative tool works. Transfer evidence only when learner, subject, setting, product and implementation are sufficiently similar.

Report results by relevant pupil group while protecting privacy. Check whether device access, reading demand, SEND support, English-language needs or prior attainment changed participation or benefit.

Protect Children’s Data Before Creating Accounts

Schools determine the purpose and roles for processing; a supplier’s privacy policy cannot make the decision for them. Map every data flow:

• identity and account information;
• prompts, pupil work and tutor conversations;
• inferred attainment, behaviour or wellbeing;
• teacher notes and safeguarding content;
• analytics, diagnostics and support access;
• model training or product improvement;
• subprocessors, storage location and transfers;
• retention, deletion, correction and export; and
• what remains when the contract ends.

DfE’s school AI data-protection guidance, updated July 2026 says use must be planned and schools should involve their data-protection or IT leads. It links AI use to DfE policy, product safety and safeguarding guidance.

The ICO’s June 2026 EdTech audit findings identified recurring gaps in controller/processor roles, contracts, data-flow maps, minimisation, retention, privacy information and DPIAs. Convert each finding into a procurement test.

Use the minimum pupil data. Never paste safeguarding files, special-category information or identifiable work into a consumer tool simply because it is convenient. Provide age-appropriate privacy information and a workable alternative when the proposed processing is optional.

Our UK AI privacy guide covers lawful basis, DPIAs and supplier controls in more detail.

Keep Safeguarding Human and Immediate

Pupil-facing systems may receive disclosures about abuse, self-harm, bullying or danger. A generic content filter is not a safeguarding response.

Before launch, define:

• which signals the supplier detects and what it does not detect;
• whether staff receive the original words, an alert or both;
• who monitors alerts during school hours and outside them;
• maximum response times and backup contacts;
• the school’s existing designated-safeguarding-lead process;
• what the pupil is told at the moment of concern;
• how false alerts and missed concerns are reviewed; and
• the manual route when the service is unavailable.

Test the end-to-end route with realistic but synthetic scenarios. Do not let a chatbot improvise crisis counselling, promise confidentiality it cannot provide or tell a child that an alert has been handled when no responsible person has acknowledged it.

Wellbeing tools require still stronger boundaries; see our AI mental-health guide.

Design Accessibility With Pupils, Not for an Average User

AI can provide text-to-speech, captions, simplified layouts, language support and alternative examples. It can also misrecognise speech, remove subject vocabulary or create easier work instead of accessible access to the same learning.

Test keyboard operation, screen readers, captions, colour, zoom, reading order, response timeouts and low-bandwidth fallback. Let learners control modality and pace. Preserve the teacher’s ability to make an individual reasonable adjustment without the system relabelling it as lower attainment.

Do not use disability or inferred need to narrow future curriculum opportunities. Record what adjustment was selected, by whom, for what task, and whether the pupil says it helped.

Our AI accessibility and inclusion guide provides a wider product checklist.

Redesign Assessment Rather Than Playing Detector Roulette

AI-output detectors are not proof of authorship. False accusations can harm pupils, while minor rewriting can evade detection. Build assessment validity around observable learning.

Ofqual’s March 2026 coursework resources advise schools and colleges in England to use a consistent whole-setting approach and explain that undisclosed AI-generated coursework is cheating. The 2026 Ofqual guide also states that AI must not be the sole marker for regulated qualifications.

For each task, tell pupils:

• whether AI is prohibited, permitted for named stages or required;
• what acknowledgement and evidence are needed;
• which process artefacts to retain;
• how teachers will authenticate work;
• how accessibility arrangements interact with the rule; and
• the review and appeal route if concern arises.

Use drafts, notes, source discussions, short oral checks, in-class components and teacher knowledge of the pupil proportionately. Investigate evidence, not a detector score.

A Six-Week Pilot That Can Stop

Suppose a Year 8 science team pilots an AI practice tutor for one unit. Teachers load approved explanations and questions. Pupils receive hints but cannot request completed homework answers. The school has completed its data map, safeguarding drill and accessibility tests.

At week two, logs show that the tutor repeatedly accepts a scientifically vague explanation. Teachers pause that item family, correct the source and regression-test it. No pupil is penalised for the system’s error. At week six, the school compares an independent curriculum assessment, misconception rate, teacher workload and participation with the baseline.

The pilot succeeds only if learning and access improve within the agreed safety bounds—not because pupils produced many chats.

Measurable Release Gates

Track teacher review minutes, substantive corrections, pupil help requests, unresolved misconceptions, safeguarding response, deletion fulfilment, accessibility failures, assessment disputes and outcomes by group. Publish limitations alongside benefits.

Personalisation Is a Teaching Decision

Education AI is valuable when it helps a teacher see where practice is needed, gives pupils another carefully bounded route into an idea, or removes low-value drafting. It is dangerous when confident text becomes curriculum, surveillance becomes personalisation or a score replaces professional judgment.

Start with a learning contract and an outcome the tool can fail. Keep the teacher in control, children’s data minimised, assessment rules explicit and human help close. That is a more demanding standard than “available 24/7”—and a far better definition of educational value.