In England, the DfE has moved from guidance to standards: its generative AI product safety expectations became a set of 13 standards, with the current version published on 19 January 2026, and KCSIE 2026 treats AI-enabled harms as a mainstream safeguarding concern.
What changed, and when
| Date | Development |
|---|---|
| January 2025 | DfE published generative AI product safety expectations for suppliers |
| September 2025 | KCSIE 2025 signposted schools to those expectations via filtering and monitoring |
| 19 January 2026 | DfE generative AI product safety standards published — 13 standards; standards define minimum requirements rather than recommended practice |
| 2026 | DfE filtering and monitoring standards updated to address generative AI content |
| 7 July 2026 | KCSIE 2026 published, in force 1 September 2026; settings work to the 2025 edition until 31 August 2026 |
The shift from expectations to standards matters. Guidance suggests; standards set a minimum. For a school procuring an AI tool, the 13 standards are now the most useful checklist available — see DfE generative AI standards.
Where AI genuinely helps in UK schools
- Lesson planning and resource adaptation. Drafting against a curriculum objective, adapting for pupils with SEND, generating retrieval practice. Teacher reviews and owns.
- Marking and feedback. First-pass scoring against a rubric with the teacher moderating. Usually the fastest measurable time recovery.
- Report writing. Drafting comments from the pupil’s actual data rather than a comment bank, which is both faster and better.
- Pattern detection. Attendance drift, participation decline, assessment divergence — surfacing what a human should look at.
- Family communication. Drafting, translating and routing across channels.
These are all adult-facing. None requires a device per pupil, and none puts a model in unsupervised contact with a child — which is why they are also the easiest to defend in a DPIA.
Where the evidence is thinner than the marketing
- AI tutoring as a substitute for teaching. Effects depend heavily on supervision. A tool that works in a supervised session frequently does nothing when set as homework.
- Automated essay marking at high stakes. Adequate for formative feedback; not reliable enough to certify, and performance varies by pupil writing style in ways that raise equity questions.
- Predictive risk models. A model trained on who was previously flagged will flag the same kinds of pupils. Useful as a prompt to look, dangerous as a verdict.
- AI detection tools. False positive rates remain high enough that they should never be the sole basis of an academic misconduct finding.
Five risks a UK school must manage
1. Pupil data leaving the school
The first failure is almost never the procured system. It is a member of staff pasting pupil names, SEND details or safeguarding notes into a consumer chatbot with no data processing agreement behind it. The school remains the controller regardless. Policy and a sanctioned alternative must arrive together — a ban without an approved tool produces hidden use, not compliance.
2. Filtering and monitoring that cannot see AI content
The 2026 DfE standards update asks schools to consider whether their filtering and monitoring can handle real-time, dynamic, personalised and AI-generated content. DNS and URL-based filters cannot see inside AI-generated material that never sits on a blocklisted page. If pupils use AI tools — and they do — this needs specific attention in the annual review.
3. AI-generated image abuse
KCSIE 2026 reframes language around image-based abuse to explicitly include AI-generated content such as deepfakes, closing a grey area many schools were stuck in. See KCSIE and safeguarding.
4. Assessment integrity
Any work completed outside supervision no longer reliably evidences the pupil. This is an assessment design problem before it is a detection problem.
5. Deskilling
If an early career teacher never plans a lesson unaided, planning expertise does not develop. Systems should make pedagogical reasoning visible rather than hiding it behind a generate button.
An adoption sequence that survives scrutiny
| Stage | Focus | Typical duration |
|---|---|---|
| 1 | Acceptable use policy, DPIA, staff briefing, governor sign-off | 4–6 weeks |
| 2 | Filtering and monitoring review specifically covering AI content | Alongside the annual review |
| 3 | Administrative load: reporting, communication, data analysis | One term |
| 4 | Teacher workload: planning, marking, adaptation | One to two terms |
| 5 | Assessment redesign for tasks a model can complete | One academic year |
| 6 | Pupil-facing tools, supervised, age-appropriate, against the 13 standards | Ongoing |
Schools that invert this — starting with pupil-facing AI because it demonstrates well — typically spend the following year retrofitting governance under pressure.
Staff training
DfE has published support materials and staff modules covering understanding AI, interacting with generative AI, safe use and practical use cases, updated during 2026. KCSIE 2026 references DfE-partnered resources covering safeguarding, ethics, data protection and intellectual property risk. The practical reading is that a school using AI is expected to understand the implications and train staff accordingly — and that DfE modules are a better basis than a vendor webinar, including ours.
Questions to ask any AI education vendor
- How do you answer each of the 13 DfE generative AI product safety standards? In writing.
- Can your product generate free-form content, or is it closed-loop and limited to approved material?
- Is pupil data used to train models? Show me the contract clause.
- Where is data hosted and processed, and who are your sub-processors?
- What logging exists, and can we see it?
- What does the tool refuse to do, and how do you know?
- What happens to our data if we leave, and what does export cost?