AI does not give a school in England a separate safeguarding duty. Keeping children safe in education 2026, in force since September 1, 2026, writes AI into duties you already hold: the definition of nudes and semi-nudes, two of the four online risk categories, and the yearly filtering and monitoring review. The job is to update four existing systems, not to write a new policy.
Sources last checked on November 13, 2026. This article covers England. Guidance and regulation in this area are moving, so each claim below carries the date of the document it comes from.
Much of what leaders are told about AI and safeguarding comes from companies that sell filtering software. The primary documents are shorter and plainer than the commentary on them, and they point to work most schools can finish in a term.
What does KCSIE 2026 actually say about AI?
It names AI in a handful of places, and every one of them attaches AI to something a school already does. I re-read the Summary and Parts one and two of Keeping children safe in education 2026 (paragraphs 1 to 209) on November 13, 2026. Part one, the section every member of staff must read, contains no AI wording at all. The references sit in the Summary and in Part two, which is written for governing bodies and leaders.
| Where | What the guidance says | What it means in your school |
|---|---|---|
| Summary, "About this guidance" (page 6, unnumbered) | Nudes and semi-nudes "may also be digitally altered or wholly generated using artificial intelligence, including what are sometimes described as 'deepfakes' or 'deep nudes'" | A fabricated image of a pupil gets your full child protection response |
| Paragraph 160, what to teach | Online harms to teach about include "the prevalence of deepfakes" | Check where this sits in your curriculum map |
| Paragraph 165, contact risk | "harmful online interaction with other users or generative AI applications that simulate this" | A chatbot conversation is now inside the online risks you plan for |
| Paragraph 165, conduct risk | "making, sending and receiving explicit images, including those generated using AI" | Your online safety policy should name AI-generated images |
| Paragraphs 167 and 168 | Points schools to the Department for Education's generative AI guidance and its staff training materials | Module 3 of those materials is your ready-made staff briefing |
| Paragraph 173 | A review of filtering and monitoring "is carried out at least once every academic year" by the responsible senior leader, with the safeguarding lead and IT support, and "a record should be kept" | The review needs a date and a minute |
| Paragraph 177 | Points to "Generative AI: product safety expectations" for how filtering and monitoring apply to generative AI | The buying standard is tied to your filtering duty |
| Paragraph 193, child protection policy | Child-on-child abuse includes making or sharing nudes and semi-nudes, "including those generated using AI" | Your policy's list should use the same words |
Two limits on that table. I did not re-read Parts three to five or the annexes today, so there may be further references there. And I have not compared every line with the 2025 edition: the law firms Farrer & Co and HCR Law both identify the nudes definition and the AI wording in the contact and conduct risks as new this year, and Farrer notes that the 2025 edition asked for the filtering review only "regularly."
The definition is the change with the sharpest edge. The Department's consultation response, published in July 2026, says it "included the term 'making' to cover different situations, such as images taken by others, images created by a child, and AI-generated content." The moment a fabricated image of one of your pupils appears, your existing process applies to it in full: the same response, the same referral thresholds, the same records, the same people.
Does a school need a separate AI safeguarding policy?
No, and nothing in the text I read asks for one. Every AI reference in the table lands on a document or routine you already have: the child protection policy, the online safety policy, the curriculum, staff training, the filtering review. A separate AI safeguarding policy duplicates those and is read by fewer people. If you want the wider policy question settled first, what a school AI policy should contain covers where safeguarding sits inside it.
Which AI safeguarding documents are law, and which are advice?
One of the four documents leaders are pointed to is statutory guidance, and the most detailed one is written for suppliers. Knowing which is which tells you where your effort has legal weight and where it only has weight if you supply it.
| Document | Status | Last updated | What it asks |
|---|---|---|---|
| Keeping children safe in education 2026 | Statutory guidance. Schools and colleges must have regard to it | In force September 1, 2026 | The eight points in the table above |
| Filtering and monitoring standards | The Department's "recommended approach" to meeting the statutory expectation | September 16, 2026 | A yearly review that covers generative AI, and a fresh review when a new AI tool arrives |
| Generative AI: product safety standards | Guidance "for edtech developers and suppliers." Schools "may also find them helpful" | January 19, 2026 | That an AI product alerts a named adult and does not manipulate a child |
| Ofcom, AI chatbots and online regulation | A regulator's account of what the Online Safety Act covers | December 18, 2025 | Nothing of schools. It describes which chatbots carry duties |
A note on names, because it trips people up. The product document was first published in January 2025 as "product safety expectations," and the statutory guidance still calls it that at paragraph 177. Since the January 2026 update its title is "product safety standards." They are the same document.
What the filtering and monitoring standards now say about AI
They bring AI-generated content inside the definition of what filtering is for. As updated on September 16, 2026, the standards give the senior leadership team the job of "scoping the specific needs of their school or college, including use of generative AI," and say that schools "introducing generative AI (artificial intelligence) products should assess the risks." The yearly review should consider whether your system "can monitor and filter real-time, dynamic, personalised or AI (artificial intelligence)-generated content." A review is also due when "new technology is introduced, such as new devices or generative-AI (artificial intelligence) tools," and it "should also involve the responsible governor."
What the product safety standards ask of an AI tool
They describe an AI product that behaves like a responsible adult would. A product should "identify and alert local supervisors to searches for, or access to, harmful or inappropriate content" and "identify and alert local supervisors of disclosures that indicate a possible safeguarding issue." Since the January 2026 update, "products should detect signs of learner distress," suppliers are expected "not to anthropomorphise products," and a product should avoid replies such as "You can trust me" and should not motivate a learner through "sycophancy and flattery."
Those are safeguarding requirements written for the tools children now use. Nothing on the page gives them legal force over a vendor, and they are addressed to suppliers, not to you. They protect children in your school only if you make them a condition of approving AI tools for schools. A school's hold over a vendor is the decision to buy, and it is worth using before the contract is signed.
Are the chatbots children use covered by the Online Safety Act?
Many one-to-one companion chatbots were outside it on Ofcom's own account, and I could not confirm today that this has changed. Ofcom's guidance on AI chatbots and online regulation, published on December 18, 2025, says "chatbots are not subject to regulation if they" meet three conditions: they "only allow people to interact with the chatbot itself and no other users," they "do not search multiple websites or databases when giving responses to users," and they "cannot generate pornographic content."
Picture the product that description fits. A one-to-one companion app. No other users. No search. No explicit content. A thirteen-year-old talking to it at one in the morning about something they have told nobody else.
The government has said it will close the gap. The law firm Osborne Clarke reported on February 26, 2026 that "the government announced plans to close the current 'legal loophole' in the OSA regarding AI chatbots" through amendments to the Crime and Policing Bill. On November 13, 2026 I could not find a primary source confirming that any such change is in force, so check the current position before you brief governors, and do not tell parents that these apps are regulated unless you have seen where it says so.
For this term, the practical position is this. The statutory guidance covers the image and names the chatbot as a contact risk. The product standards cover the tool you chose to buy. The app a child downloaded on their own phone may be covered by neither, and whatever protection that child has in that conversation is the protection your school built. If that is the case in front of you, what a principal should do when students confide in AI chatbots takes it further.
The Absorption Test: four checks on systems you already run
The work AI creates for you is four checks, not a new policy. That is the question I put to leadership teams who tell me they need an AI safeguarding policy, and I call it the Absorption Test: does the safeguarding machinery you already run absorb AI, or does AI fall straight through it? Four systems are tested, in this order: the definitions in your child protection policy, your filtering and monitoring, what your staff and your designated safeguarding lead actually know, and the route a concern travels once someone raises it. A school that passes has changed four documents and one training slide. A school that fails writes a new policy nobody reads, and the concern still arrives at the IT helpdesk.
I offer this as a way of framing the work, not as a tested method. Take the four checks in order, because each one depends on the one before it.
Does your policy's definition include images that were never real?
Open your child protection policy and find where it defines a nude or semi-nude image. If that definition describes a photograph, it is now narrower than the statutory guidance, and that sentence is the one a member of staff will read while deciding whether what a pupil has shown them counts. Widen it to match the Summary definition, including digitally altered and wholly AI-generated images, and make the child-on-child abuse list match paragraph 193. Then check the same wording in your acceptable use policy, your behavior policy and your online safety policy, because it usually appears in all four and is usually corrected in one.
Can your filtering and monitoring see AI tools at all?
Your system may have been configured against categories that predate generative AI, so ask your provider two questions: which AI categories exist, and what happens to a site that fits none of them. Then ask the question the September 2026 standards now pose, which is whether the system can deal with content generated in real time, not only with pages already on a blocklist. Confirm the yearly review is happening, led by the senior leader responsible, with the safeguarding lead, IT support and the responsible governor involved, and a dated record at the end. A review nobody can date did not happen.
Do your staff know this is a safeguarding matter and not an IT matter?
This is where I see schools fail, and it fails quietly. A hypothetical case, but a familiar shape: a pupil tells a teaching assistant that someone has made a fake image of them. The assistant, acting in good faith, treats it as a technology problem, because the word AI was in the sentence. It goes to the network manager. Two days pass. Under the statutory definition that was a nude image of a child, and the clock started when the pupil spoke.
One slide in your next safeguarding briefing fixes it: if it involves a child, it goes to the designated safeguarding lead, whatever the technology in the story. The same slide should say that a child confiding in a chatbot is a contact risk under paragraph 165, and that a disclosure made to a machine leaves no trail anyone in your building can see. The Department's own Module 3 on the safe use of generative AI, which paragraph 168 points to, gives you the rest of the briefing.
If you hold the safeguarding lead role yourself, what a designated safeguarding lead should change because of AI covers the changes to make to your own records and your approved tools. And if some of your staff are further ahead on these tools than you are, what to do when your team knows more about AI than you covers how to use that knowledge in the briefing.
Where does the concern go, and does anyone check that it arrived?
Follow one real route end to end. A pupil raises a fabricated image with a lunchtime supervisor. Where does it go, how fast, who records it, who decides on a referral, and who tells the pupil what happened next? Do the same for a concern raised by a parent and for one thrown up by monitoring software. If any of the three routes has a step that exists only in someone's head, that is your term's work, and it has nothing to do with AI. It is the same question as who is accountable when AI gets it wrong, asked about children.
What should a school change this term?
Ten things, each with an owner and a date, and none of them is a new policy. Copy this list into your next leadership agenda and fill in the two blanks on each line.
- Child protection policy: the definition of nudes and semi-nudes includes digitally altered and wholly AI-generated images. Owner: ____ By: ____
- Child protection policy: the child-on-child abuse list reads "making or sharing," and names AI-generated images (paragraph 193). Owner: ____ By: ____
- Acceptable use, behavior and online safety policies: the same wording carried across all three. Owner: ____ By: ____
- Online safety policy: contact risk names generative AI applications; conduct risk names AI-generated explicit images (paragraph 165). Owner: ____ By: ____
- Curriculum map: where pupils are taught about deepfakes, and in which year (paragraph 160). Owner: ____ By: ____
- Filtering and monitoring: the yearly review is done, minuted and dated, with the senior lead, safeguarding lead, IT support and responsible governor (paragraph 173). Owner: ____ By: ____
- Filtering and monitoring: the provider has answered in writing what the system does with AI-generated content and with AI sites it has not categorized. Owner: ____ By: ____
- Tool approval: no pupil-facing AI tool is approved until the supplier has shown how it alerts your safeguarding lead to a disclosure. Owner: ____ By: ____
- Staff briefing: one slide, one sentence. If it involves a child, it goes to the safeguarding lead, whatever the technology. Owner: ____ By: ____
- Reporting routes: one concern walked end to end from a pupil, a parent and the monitoring system, with the breaks written down. Owner: ____ By: ____
Who sits in the room when those owners are named is the question who should sit on your school's AI steering group answers.
What I Tell Leadership Teams
That the volume of AI safeguarding advice is inversely related to its usefulness. Across the leadership teams I work with, the pattern is consistent: the schools that feel least prepared are usually the ones that have read the most, because much of what is available is written by companies whose product is the answer to the question they have just posed. I have spent time on the other side of that conversation, advising the Department for Education, KHDA in Dubai and the Ministry of Education in Kazakhstan on this sort of guidance, and as a school trustee reading the safeguarding report, not writing it. The documents that survive contact with a real school are short, and they change something that was already there.
The other thing I say lands less comfortably. If your AI safeguarding response is a new policy, you have chosen the artifact that is easiest to produce and hardest to use. Nobody reads the second safeguarding policy. They read the first one, once, in September, and then they act on what they remember. Put the change where they will meet it. When a trust or a conference wants that argument put to a whole room, it is one of the things I cover in my education keynotes.
Is the concern about AI and safeguarding overblown?
For the average teenager, partly; for the children safeguarding exists for, no. Pew Research Center's survey of US teens aged 13 to 17, conducted between September 25 and October 9, 2025 and published in February 2026, found them using chatbots mostly for ordinary things: 57 percent to search for information, 54 percent for help with schoolwork, 47 percent for fun. Only 12 percent had used one for emotional support or advice.
Safeguarding has never operated on the average child. Internet Matters, surveying 1,000 UK children aged 9 to 17 for its report Me, Myself and AI, published on July 13, 2025, found 64 percent of children using chatbots and 35 percent of those users saying it is like talking to a friend. Among the children the study classed as vulnerable, 71 percent used chatbots. Of those vulnerable users, 26 percent said they would rather talk to a chatbot than to a real person, and 23 percent said they use one because they have nobody else to talk to. That is the group your safeguarding system exists for.
The image risk is not hypothetical either. The Internet Watch Foundation reported on August 10, 2026 that its Report Remove helpline, run with Childline, received 420 reports from British under-18s in the first six months of 2026 involving imagery the child believed was faked, against 397 across the whole of 2025. By the end of June 2026, 268 actioned reports involved suspected faked imagery, compared with 221 in the whole of 2025. Half a year has passed a full year. Some of those children are sitting in registration on Monday.
What good looks like by the end of this term
A school that has taken this seriously has four dated records and no new policy. The definition in the child protection policy matches the statutory one, and the same wording has been carried into the other documents that repeat it. The filtering and monitoring review has happened, is minuted, and records what the system does with AI-generated content. Every member of staff has been told, in one sentence they can recall under pressure, that anything involving a child goes to the safeguarding lead whatever the technology. And one reporting route has been walked from end to end by someone who then wrote down where it broke.
None of that requires a view on whether AI is good for education. It requires you to notice that the guidance changed, and to move a few pieces of paper before somebody needs them.
Three related questions sit close to this one. If you are new in post, the first 90 days of leading AI change sets out where these checks sit among everything else arriving in the same fortnight, and the seven dimensions of AI readiness show whether the staff knowledge and reporting routes behind them are strong enough to hold. If the data question is the one in front of you, start with whether staff can put student data into ChatGPT. If a pupil is already in the frame, what to do when a student is accused of using AI sets out the process. And if an inspection is on your mind, what Ofsted expects from schools on AI covers how this is likely to be looked at.
Governance work of this kind is rarely urgent until it is, and the schools that come through an incident well are the ones that did the boring version of it in a quiet week. AI training for schools is where I help leadership teams turn a guidance change into something their staff can act on, and my newsletter carries the guidance changes as they land.
Sources and further reading
- Department for Education, Keeping children safe in education 2026, statutory guidance, published July 7, 2026, in force September 1, 2026. Summary and paragraphs 160, 165, 167, 168, 173, 177 and 193 read on November 13, 2026.
- Department for Education, Keeping children safe in education 2026: government consultation response, July 2026.
- Department for Education, Filtering and monitoring standards for schools and colleges, last updated September 16, 2026.
- Department for Education, Generative AI: product safety standards, first published January 22, 2025 as "product safety expectations," last updated January 19, 2026.
- Department for Education, Safe use of generative AI in education: module 3, last updated May 19, 2026.
- Ofcom, AI chatbots and online regulation: what you need to know, December 18, 2025.
- Osborne Clarke, Regulatory Outlook: digital regulation, February 26, 2026.
- Farrer & Co, Keeping children safe in education 2026: a practical guide for schools, 2026.
- HCR Law, KCSIE 2026: key safeguarding updates, July 22, 2026.
- Internet Matters, Me, Myself and AI: understanding and safeguarding children's use of AI chatbots, July 13, 2025 (1,000 UK children aged 9 to 17).
- Pew Research Center, How Teens Use and View AI, February 2026 (US teens aged 13 to 17, fieldwork September 25 to October 9, 2025).
- Internet Watch Foundation, Hundreds of British under-18s reporting nude or sexual imagery of themselves say it's faked or manipulated, August 10, 2026.
Change log. November 13, 2026: rebuilt against the primary documents. An earlier version said the statutory guidance made one explicit change about AI, in "paragraph 6." That was wrong on both counts: the definition sits on page 6 of the Summary, and the guidance names AI in at least seven other places, now set out in the first table. The September 2026 filtering and monitoring standards and a ten-point term list were added.
Dan Fitzpatrick is the founder of The AI Educator and writes and speaks on how leaders should respond to AI. More about Dan.


