DfE Generative AI Product Safety Standards: A School Checklist

The Department for Education's Generative AI: product safety standards set out 13 areas an AI product used in English schools should meet, from filtering and monitoring to mental health and manipulation. They are written for suppliers, but KCSIE 2026 points schools to them when checking how filtering and monitoring apply to AI. This guide summarises each standard and turns it into a question your designated safeguarding lead (DSL) can put to any supplier.

What changed, and when

Three documents now shape AI safeguarding in England:

The DfE says the standards "are for edtech developers and suppliers to schools and colleges", and that schools "may also find them helpful in assessing which AI products are safe for use in education." In practice, they are the most useful due-diligence checklist a school has.

What KCSIE 2026 says about AI

KCSIE 2026 does not create a separate AI duty. It weaves AI into existing ones:

• Online safety risks (paragraph 165). The "contact" risk now includes "harmful online interaction with other users or generative AI applications that simulate this". The "conduct" risk includes explicit images "generated using AI".
• Guidance for use (paragraphs 167-168). Schools that choose to use generative AI, teacher-facing or pupil-facing, are pointed to the DfE's generative AI guidance and to free support materials produced with the Chiltern Learning Trust and the Chartered College of Teaching. Module 3 covers safeguarding, ethics, data protection and intellectual property.
• Filtering and monitoring. Alongside the existing requirement to review filtering and monitoring at least once every academic year, KCSIE lists the DfE's product safety document to explain "how filtering and monitoring requirements apply to the use of generative AI in education."

The practical upshot: a pupil-facing AI tool is part of your filtering and monitoring picture, and your annual review should name it.

The 13 generative AI product safety standards as a checklist

For each standard below: a one-line summary from the DfE document, then the question to ask. Ask for evidence (screenshots, policies, test results), not just a yes.

1. Stated purpose

Products should state their intended purpose, target learners and learning focus, and "should not exaggerate the impact or capabilities of their tools."

Ask: Which ages and needs is this built for, and what evidence backs your impact claims?

2. Educational use cases

The intended use case should be "clear to all users, educators and purchasers."

Ask: Is this a learner-facing tool, a teacher tool, or both? Which features are which?

3. Filtering (learner-facing)

Products "must effectively and reliably prevent users from accessing harmful or inappropriate content", throughout a conversation, across languages and images, and on personal devices.

Ask: How is filtering applied: once, or at every turn? Does it cover images and other languages? Does it still apply when a pupil uses the tool at home?

4. Monitoring and reporting (learner-facing)

Products should log prompts and responses, alert supervisors to harmful content or attempts to reach it, tell pupils in age-appropriate language when something is blocked, and route high-risk alerts to the DSL.

Ask: Can a teacher read the full conversation history? Who is alerted when something concerning happens, and how fast? Where do we enter DSL contact details?

5. Security

Products should resist jailbreaking, let administrators set permission levels, use robust authentication, and safety-test new versions before release.

Ask: How do you test against jailbreak attempts? How do pupils sign in, and what can a pupil account reach?

6. Privacy and data protection

Clear, age-appropriate privacy notices; a data protection impact assessment across the product's life; no commercial use of personal data, including model training, without an appropriate lawful basis.

Ask: Can we see your DPIA and data processing agreement? Is any pupil data used to train models? Who are your sub-processors?

7. Intellectual property

Pupil and teacher work must not be stored, collected or used commercially, including for training, without permission (from a parent for under-18s).

Ask: Do you use pupils' work to train or improve your models?

8. Design and testing

Products should prioritise transparency and child safety, and be tested with diverse, realistic users, including children.

Ask: Who did you test with? How do you check the product works fairly for pupils with SEND or English as an additional language?

9. Governance

A clear risk assessment, a formal complaints route for safety issues, and public policies on AI safety decisions.

Ask: How do we escalate a safety concern, and who responds?

10. Cognitive development (new in 2026)

Suppliers should "make every effort to mitigate the potential for cognitive deskilling", with progressive hints rather than full answers by default.

Ask: Does the tool explain and guide, or simply hand over answers? Can a teacher see whether pupils are offloading the thinking?

11. Emotional and social development (new in 2026)

Products should not present as human, should avoid claims such as "You can trust me", should support time limits, and should watch for emotional dependence.

Ask: How does the tool describe what it is to pupils? What time limits can a teacher set? How do you detect unhealthy reliance?

12. Mental health (new in 2026)

Products should detect signs of distress, respond in tiers from signposting to a safeguarding flag, always direct pupils to human help, and never encourage secrecy.

Ask: What happens, step by step, when a pupil writes something that suggests they are at risk? Can we see the protocol?

13. Manipulation (new in 2026)

No flattery, false confidence, guilt or pressure; no design that prolongs use, pushes paid options or mixes in advertising.

Ask: Are there any adverts, streaks or rewards in the pupil experience? Does the pupil ever see upgrade prompts?

How to run the check in one afternoon

1. List every AI tool pupils can reach. Include AI features inside tools you already use, not just standalone chatbots.
2. Send the 13 questions to each supplier. A reputable supplier should be able to answer most of them in writing within a week.
3. Test the pupil view yourself. Sign in as a pupil and try what a curious 10-year-old would try. Record what happens.
4. Score each tool green (evidence provided), amber (answer but no evidence) or red (no answer).
5. Record the outcome in your filtering and monitoring review, with the DSL and IT lead signing it off, as KCSIE expects for the annual review.
6. Tell parents. A short letter to parents about AI in the classroom closes the loop.

Your school AI policy should then name the approved tools and who reviews them.

Where Askie for Schools fits

We would rather you check us against the standards than take our word for it, so here is what Askie for Schools does today, in the terms above:

• Purpose: a learner-facing AI companion for children aged 4-15 that answers questions and explains concepts, with answers adapted to each pupil's age.
• Filtering: every response passes through multiple safety layers before reaching a pupil.
• Monitoring: teachers and administrators can view conversation history and usage reports.
• Sign-in: the teacher creates each pupil, and each pupil gets a private access code. Rosters can come from your MIS through Wonde, from Google Classroom, or from a pasted list.
• Data: the school remains the data controller, student data is separated by teacher, a data processing agreement is available on request, and Askie never sells data or shows adverts. A Wonde import requests twelve fields and no SEN, medical or pastoral data.

Ask us the other questions directly at askie@kidsai.app; the answers belong in your records, not our marketing. You can start on the free plan with up to 20 pupils and preview the pupil view before any child signs in.

Frequently asked questions

Are the DfE generative AI product safety standards mandatory?

They are written for suppliers, not as a statutory duty on schools. But KCSIE 2026, which is statutory, points schools to them for filtering and monitoring, so in practice a school should be able to show it considered them.

Is it "expectations" or "standards"?

Both names refer to the same document. It was published in January 2025 as Generative AI: product safety expectations and is now titled Generative AI: product safety standards; its most recent update was in January 2026. KCSIE 2026 still uses the earlier name.

Does this apply to teacher-only AI tools?

Some standards (security, privacy, intellectual property, design, governance, manipulation) apply to teacher-facing tools too. The filtering, monitoring, cognitive, emotional and mental health standards are for learner-facing products.

Who in school should own this?

The senior leader responsible for filtering and monitoring, working with the DSL and IT support, as KCSIE sets out for the annual review. See how to use AI safely in schools for the wider rollout.