The first job with AI in schools is to find out how your staff and pupils already use it. Then put the safeguarding and data checks in place before you buy or ban anything. Only after that should you write the policy, set the assessment rules and train your staff.
Say you’re a secondary head, and a parent emails on a Monday. She wants to know if pupils can use chatbots for homework. You go looking and find three answers. The IT lead thinks the big ones are blocked. Your head of English lets Year 10 use them for first drafts, and the exams officer’s got the JCQ poster up by the photocopier. Nobody’s done anything wrong. The tools just got into the building before the rules did.
In short, and in the order I’d take it:
- Find out what’s already happening, from staff and pupils and from the platforms you already pay for.
- Put the safeguarding and data floor in.
- Write a short policy that people will read, and review it every term if you can, and at least once a year.
- Settle the rules for assessed work.
- Point AI at workload, and train staff before you ask them to use it.
- Teach pupils to question it, and leave the big purchases until the rest of this is in place.
What AI in schools looks like now, and where to start
AI’s already in your school. It’s used by more staff than you’d guess, with fewer rules than you’d want. So the first move is a proper look at what’s happening.
The Department for Education’s technology survey asked 1,211 teachers in England about it between January and May 2025. In total, 44% used generative AI at least sometimes for one or more teaching tasks (DfE technology in schools survey, 2025). Lesson planning was the commonest use. In secondary schools, 73% of teachers thought their pupils had used it for homework in the previous month.
Later figures run much higher. In November and December 2025, only 12% of teachers in England told Teacher Tapp they’d never used AI for schoolwork (Teacher Tapp in Impact, 2026). It’s a different question, so treat the two as separate snapshots.
Policies lag further behind. In the DfE’s 2025 survey, just 22% of IT leads said their school had a policy on the safe and appropriate use of generative AI. Another 23% didn’t know if it had one.
Ofsted’s 2025 study of early adopters shows the same picture from the keen end. It’s a small study, of leaders at just 21 schools and colleges that started early, so it isn’t typical. Most had an AI champion who brought senior leaders on board, and leaders tended to take small first steps before going wider (Ofsted early adopters report, 2025). Very few had a longer-term plan beyond the pilot.
So start with an audit. You don’t need to invent one. The DfE’s leadership toolkit on AI, updated in May 2026, has one built in (DfE leaders’ AI toolkit, 2026). It covers where you’re at now on strategy, policies, staff training, curriculum, pupil use and procurement.
I’d add pupils and parents to it. Staff will tell you what they’ve tried. Pupils, asked without their names on it, are the ones who’ll tell you what’s really going on. A handful of parents can fill in the kitchen-table view. If the staff room splits into ‘ban it’ and ‘buy everything’, the audit gives both camps the same facts.
Put the safeguarding and data floor in first
Before pupils use a new AI tool, check that it’s covered by your data protection as well as your filtering and monitoring. That’s the floor everything else stands on. A lot of it’s already expected of you.
Keeping children safe in education 2026 has been in force since 1 September 2026, and it names generative AI in its online safety risks (KCSIE 2026). Its ‘contact’ risk includes generative AI applications that simulate harmful interaction with other users. Under ‘conduct’, it lists explicit images generated using AI.
The 2026 version also updates the wording on filtering and monitoring. Governing bodies and proprietors should make sure those systems are reviewed at least once every academic year. The senior leader responsible should lead it, with the DSL and IT support, and keep a record of the checks.
I’d sit down with your DSL and, line by line, go through what KCSIE 2026 changes about AI.
The DfE’s guidance on generative AI covers the data side (DfE generative AI guidance, 2025). It recommends keeping personal data out of these tools. Schools mustn’t let pupils’ original work be used to train them, unless they’ve got permission or a copyright exception applies. You’re also expected to keep to the age restrictions AI tools set. UNESCO’s 2023 guidance went further and proposed an age limit on independent conversations with generative AI platforms (UNESCO generative AI guidance, 2023).
Or, as I’ve put it before: “You wouldn’t give a 10-year-old the keys to a sports car who’s never driven it and say, ‘Go on, son, knock yourself out.’ That’s what we’re doing with AI.”
For any tool that touches pupil data, do a data protection impact assessment with your data protection officer before it’s switched on. The DfE toolkit recommends one before you use new AI tools. And the ICO says a DPIA is required where processing is likely to result in a high risk (ICO DPIA guidance, 2025). New technology and processing that targets children are both on its list of warning signs.
Then ask every supplier the same set of questions. That includes the AI features added to platforms you’ve already got.
The DfE first published its generative AI product safety standards in January 2025 (DfE product safety standards, first published 2025). They’re written mainly for suppliers, but the DfE says schools may find them helpful in judging which products are safe.
A January 2026 update added four more: cognitive development, emotional and social development, mental health and manipulation. So the list now runs well past filtering and data protection. They’re a good spine for your own questions, and these are the six I’d ask before any pupil logs in.
| Ask the supplier | A good answer | Where the rule comes from |
|---|---|---|
| What problem does it solve for us, and what’s the evidence? | A named task, like report writing, and results from schools like yours | DfE toolkit |
| Is our data, or pupils’ work, used to train your model? | No, and it’s in the contract | DfE guidance |
| What does it filter, and what can staff see? | Harmful content blocked, with logs your DSL can check | KCSIE 2026 |
| What age is it built for? | A minimum age that fits the pupils using it | DfE guidance |
| Which DfE product safety standards do you meet? | Each one named, with evidence | DfE product safety standards |
| Will you help with our DPIA? | Yes, with a map of where the data goes | DfE toolkit and ICO |
These are the rules for schools in England as I checked them on 9 October 2026, though the ICO’s data protection rules apply across the UK. It’s general information, not legal advice, so check the detail with your DSL, your data protection officer or a qualified adviser.
Write a short AI policy, with the assessment rules in it
A policy people read beats a long one nobody opens. So keep yours to a page or two and write it for the staff room. Review it every term if you can, and at least once a year.
There’s no single right format. Some of Ofsted’s early adopters wrote a separate AI policy, and others added AI to policies they already had, like safeguarding and acceptable use. Several reviewed their AI policies at least termly. The DfE toolkit leaves you the same choice.
Either works. What doesn’t is guidance that lives in one person’s head. In the DfE survey’s interviews, leaders and IT leads more often described informal guidance than a formal policy.
Five things go on the page. Staff need to know which tools they can use and what for, and pupils need the same. Everyone needs to know what never goes into a tool, and who checks AI output before it’s used. And pupils need to know how to show where they’ve used it. Get those five right and you’ve got most of an AI policy for your school.
When pupils use AI to cheat
For assessed work, the Joint Council for Qualifications has already drawn the line. AI misuse is where a student’s used AI tools without appropriately acknowledging it (JCQ guidance on AI use in assessments, 2025). It counts as malpractice. Centres should review their malpractice and plagiarism policy for AI regularly, and tell students what appropriate and inappropriate use looks like.
The DfE survey shows why that’s worth doing. Among secondary leaders whose pupils could get to generative AI, 67% had come across plagiarism. The DfE’s guidance also suggests reviewing homework and other unsupervised study, now these tools are so easy to reach. Detection software won’t carry this on its own. Tasks where more of the thinking happens in front of you are the longer-term way to respond to AI cheating.
Point AI at workload, and train staff before you ask
Aim AI at the admin and preparation that eat teachers’ evenings, and keep it well away from the relationship with pupils. And train people before you expect them to use it. My bet is that untrained staff are the likeliest to paste in something they shouldn’t.
Where AI fits in a teacher’s week
Mostly around the lesson rather than in it. In the DfE technology survey, only 7% of teachers used it to deliver live lessons, against 35% for lesson planning. Whether it’ll replace teachers altogether is a bigger question. I’ve gone through the trial evidence in my piece on whether AI will replace teachers.
The part it can’t touch is the adult who knows the class. A pupil who’s gone quiet since half term still needs someone who’s noticed, and I’ve written about motivating students who’ve switched off.
Teacher burnout and the time AI saves
AI can take some admin off a tired staff. Unless you decide where the saved time goes, though, it won’t touch teacher burnout. In 2025, 86% of teachers and leaders in England said they experienced stress in their work (DfE working lives survey, 2025). And 57% agreed their job negatively affects their mental health. If the minutes AI saves fill straight back up with admin, nothing changes for those people.
What I’m describing is workload practice, not treatment. If someone on your staff is struggling, point them to their GP or to occupational health.
Training staff will actually use
Use has run ahead of training. In the DfE technology survey, 29% of teachers had been trained in how to use generative AI in the previous 12 months. That’s against the 44% already using it. About a fifth of primary schools (21%) and a quarter of secondaries (25%) were offering that training. Another 46% of primaries and 53% of secondaries were planning to.
The DfE’s free support materials are where I’d start (DfE AI support materials, 2026). You’ll find four modules for all staff and a set for leaders. The DfE says module 3, on safe use, should be used whatever your level of experience.
Then train on your own work, because that’s where it sticks. Bring a report someone’s got to write next week, do it with the tool in the room, and check what comes out together. That’s how I’d run AI training for teachers.
Teach pupils to question AI
Pupils need to know what these tools do, where they’ll go wrong and when to leave them shut. That’s a teaching job as much as a filtering one.
Ofsted’s early adopters put teaching pupils to use AI safely among their top priorities, and the DfE survey shows why. Pupils being exposed to misinformation came second for secondary leaders whose pupils could access these tools, at 53%. For primary leaders it came first, at 28%.
Knowing how the tool works and where it fails is the core of AI literacy. I’d pair that with the human side: the five intelligences I’d argue AI can’t automate, which are Curiosity, Creativity, Critical Thinking, Adaptive Leadership and Emotional Intelligence. It’s my own way of sorting what to protect, and it hasn’t been tested as a model. Still, it gives a staff room something concrete to plan against.
Critical Thinking is the obvious one here. Picture a Year 8 checking a chatbot’s answer against the textbook and catching the date it got wrong. That’s the habit you’re after, and I’ve set out five habits for pupils who question AI in more detail. Every one of them comes back to “Don’t outsource your thinking.”
What can wait until the floor is in
Big purchases can wait. I mean AI marking at scale, VR headsets and whole-school immersive platforms. They’ll all work better once the policy, the data checks and the training are done. Buying first is the dearest way to learn that, and it’s the case I made in stop buying AI and start building learners.
AI marking
In the DfE survey, marking was the task the fewest teachers used it for, at 5% of all teachers. I think that’s sensible while the rest of the floor goes in. The DfE’s guidance says anything AI produces needs checking with professional judgement, and the professional who produced the final work stays responsible for it. Low-stakes work is the place to pilot AI marking. Have a teacher check every piece before it goes near anything that counts.
VR and immersive learning
I’m the co-founder of an immersive learning company, and I’d still tell you VR can wait until the basics are in. Headsets bring their own checks. There’s the manufacturer’s age guidance, safeguarding in shared virtual spaces, and who charges them and who cleans them. Sort all that before you judge what VR does for learning.
When you’re ready, start with one subject and one learning goal, and judge it on that. A trolley of headsets looking for a use doesn’t help anyone, because the lesson has to come first in immersive learning.
A first-term checklist for school and trust leaders
By the end of your first term on this, I’d want each of these ticked off.
- Name one senior leader as AI lead, and tell governors or trustees who it is.
- Survey staff on which tools they use and what for, then ask pupils the same thing, anonymously this time.
- List every AI tool in use, built-in features included.
- Check that filtering and monitoring cover generative AI, and keep a record.
- Do a DPIA with your data protection officer for any tool that touches pupil data.
- Put the six questions to every vendor.
- Write a one or two-page AI policy, or fold AI into the policies you’ve got. Review it each term, or at least once a year.
- Update your malpractice policy against the JCQ guidance, and show pupils how to acknowledge AI use in their work.
- Get every member of staff through the DfE’s module on safe use.
- Time one workload task before and after a short pilot, and agree with staff, in writing, where the saved time goes.
- Book a session with governors or trustees at the end of term.
In a trust, run the first two at every school before you set anything centrally. That way the trust policy fits what’s really happening.
How Colin helps
With heads and trust leaders, my job is getting the order of these moves right. Sometimes that’s a keynote so the whole staff hears the same plan, and sometimes it’s advisory work alongside the senior team. I’d anchor it on the five human intelligences, because they keep everyone focused on what pupils still need to do for themselves.
My TEDxSwansea talk was called “Bridging the Gap: How AI and VR Can Rebuild Education in the Real World”. It connects what schools teach with what the working world now pays for, and that’s still the gap I work on.
At Illuminate XR, where I’m co-founder and co-CEO, our team builds immersive AI and VR learning for schools. That puts me on the supplier side of those six questions. It also gives me a commercial interest in how schools buy this kind of tech.
If you’d like to talk about a keynote or advisory work for your school or trust, get in touch.
Related questions
Are pupils allowed to use AI in school?
That’s up to each school. The DfE’s guidance for England leaves it with schools, on conditions. Keep to the age restrictions AI tools set, keep personal data out and keep pupils safe online. For assessed work, the JCQ rules on acknowledging AI use apply as well.
What are the disadvantages of AI in schools?
The problems secondary leaders report most are plagiarism and misinformation. In the DfE’s 2025 survey, 67% of secondary leaders whose pupils could access generative AI had come across plagiarism. And 53% knew of pupils being exposed to misinformation. A school counted if its leaders knew of even one incident, so the figures don’t tell you how often it happens.
Can teachers tell if pupils use AI?
Not reliably, from software alone. JCQ says detection tools can be part of the evidence. It also notes they give lower scores when AI-generated text has been edited, and it suggests using more than one. Knowing a pupil’s work, and seeing their drafts as they go, still tells a teacher more.
Does this apply to universities?
Partly. KCSIE is statutory guidance for schools and colleges, so a university won’t be working from it. It’ll set its own rules for AI in higher education. The order of moves still holds there, so start by finding out what’s actually happening.