AI Safety: DfE Generative AI Product Safety Standards
This statement was last updated on 24 June 2026 and explains how EVA™ (Grey Gecko Educational Technology Ltd) maps to the Department for Education’s Generative AI product safety standards, version 19 January 2026.
1. The architecture that shapes everything
EVA’s generative AI runs at the authoring stage, inside a controlled pipeline under human and expert review. What reaches a pupil is a finished, reviewed, static lesson — structured talk tasks, self-contained activities and simulations. There is no live AI model generating content in front of a class, and no free-text conversation between a pupil and a model.
That architecture is deliberate, and it is also pedagogical: EVA is built so that teacher-led and peer talk do the thinking work, not a chatbot. It means the standards aimed at live conversational AI are met by design rather than by bolting on filters and monitors after the fact. EVA’s primary classification is use case 1 — content creation and delivery, teacher-mediated.
2. How we report status
Each standard below is marked Met by design (satisfied by EVA’s architecture) or In progress (evidence or documents being finalised ahead of launch). We state status honestly: where conformance rests on a document still being prepared, we say so. The authoritative text of the standards is the gov.uk publication.
3. Standard-by-standard
1. Stated purpose — In progress
What the DfE asks: state the product’s purpose, use cases and intended age group clearly, and never overstate impact — claims must rest on transparent evidence.
EVA’s position: EVA is a content-creation-and-delivery tool for Key Stages 2–4, used by teachers to deliver dialogic, oracy-focused lessons. Evidence claims on EVA’s public site reference the dialogic and oral-language research base EVA is built on (Alexander, Mercer, EEF), and are being audited so it is unambiguous these are the foundation EVA draws on, not EVA’s own measured outcomes.
2. Filtering — Met by design
What the DfE asks: reliably prevent users reaching harmful or inappropriate content, with filtering embedded in the product and sensitive to context.
EVA’s position: there is no live generative endpoint for a pupil to prompt, so there is no real-time output to filter. Content safety is applied upstream, at authoring: every lesson passes review and quality-assurance gates, to which a dedicated safeguarding-content classification step is being added. EVA does not bypass or weaken a school’s own filtering.
3. Monitoring and reporting — Met by design
What the DfE asks: keep robust activity logs and alert designated safeguarding leads to risks such as distress or self-harm.
EVA’s position: EVA has no live pupil-AI interaction, so there is no conversation to monitor and no surface through which a pupil could disclose distress to a model. Safeguarding remains with the school’s existing monitoring stack and its designated safeguarding lead.
4. Cognitive development & over-reliance — Met by design
What the DfE asks: mitigate cognitive deskilling; favour progressive disclosure — hints and scaffolds over finished answers.
EVA’s position: EVA’s entire pedagogy is the opposite of offloading — it is designed to increase pupil reasoning and talk. No AI answers on a pupil’s behalf at runtime because no AI runs at runtime. EVA’s Dialogic Fading System deliberately scaffolds support and then withdraws it. A published child-development impact plan is in preparation.
5. Emotional & social development — Met by design
What the DfE asks: avoid anthropomorphism, personhood cues and emotional dependence; avoid names, avatars or characters that imply a person.
EVA’s position: “EVA” is a product and platform name, like a curriculum series — not a character a pupil talks to, confides in, or forms a bond with. EVA presents no companion persona and uses no first-person emotional address to pupils. EVA uses no live-generation activity with pupils; any roleplay element is pre-authored, time-limited and visibly framed as an exercise. A short anthropomorphisation policy is being documented.
6. Mental health — Met by design
What the DfE asks: recognise signs of distress, signpost appropriately, and avoid giving clinical or crisis advice.
EVA’s position: EVA cannot receive a pupil disclosure because there is no live pupil-AI exchange. It does not generate content on first aid, specific clinical or legal advice, or individual child-protection matters; sensitive topics are handled at authoring with guidance for teachers.
7. Manipulation — Met by design
What the DfE asks: no sycophancy, no dark patterns, no design that maximises engagement at the user’s expense.
EVA’s position: EVA is not an engagement product. There are no streaks, push notifications, reward loops or other retention mechanics aimed at pupils. Use is teacher-mediated and task-focused; the lesson, not screen time, is the unit of value.
8. Security — In progress
What the DfE asks: authentication, role-based permissions, prompt-patching, resistance to malicious use, and alignment with the Cyber Security Standards for Schools and Colleges.
EVA’s position: access uses authenticated accounts with permission levels appropriate to role. The absence of a live model endpoint materially reduces the prompt-injection attack surface. Formal security attestations, and independent penetration testing as EVA scales, are being scheduled.
9. Privacy and data protection — In progress
What the DfE asks: comply with UK GDPR and the Data Protection Act 2018, complete a DPIA across the lifecycle, provide an age-appropriate privacy notice, honour the ICO’s Children’s code, and make a data-processing agreement available to the school as controller.
EVA’s position: EVA is built to process little or no pupil personal data, by design, and Grey Gecko Educational Technology Ltd is registered with the ICO as a data controller. Our Privacy Notice is published; a Data Protection Impact Assessment is in preparation for sign-off by a data-protection adviser, alongside a school-counter-signable processing agreement.
10. Intellectual property — In progress
What the DfE asks: do not store or use learner or teacher IP for commercial purposes such as model training without consent; be clear on the IP status of inputs and outputs.
EVA’s position: EVA makes a binding commitment not to use teacher or pupil work to train or fine-tune any model, and does not ingest such work for that purpose. Teachers retain ownership of anything they create. Third-party source materials used within lessons are being reviewed to confirm appropriate licensing or permitted use.
11. Design and testing — In progress
What the DfE asks: prioritise transparency and children’s safety in design, test with a diverse and realistic range of users including children, and test new versions before release.
EVA’s position: EVA lessons are quality-assured before release through structured assessment gates, including a dialogic-quality instrument and an oracy gate. Founding-school pilots provide real-classroom testing evidence, which is being logged systematically against this standard.
12. Governance — In progress
What the DfE asks: operate with accountability — a clear product risk assessment and a named person responsible for AI-safety decisions.
EVA’s position: a product-level risk assessment is being finalised, with a named accountable owner — Faye Fulton, Founder — responsible for safety decisions and review.
13. Accountability & transparency — In progress
What the DfE asks: publish the policies governing AI-safety decisions, provide a complaints and escalation route, and take ongoing responsibility as models and risks evolve.
EVA’s position: this page is the public face of that commitment. A complaints and escalation route is available at faye@greygecko.ai, with a stated review cycle, and this mapping is re-verified against the DfE source on each cycle.
4. What EVA deliberately does — and does not — do
Our commitments:
- Generative AI is used only at the authoring stage, under human and expert review.
- Delivered classroom resources are static and reviewed — no live model output.
- We do not train or fine-tune any model on teacher or pupil work.
- Teachers retain ownership of what they create.
- We re-verify this mapping against the DfE source on a stated cycle.
What EVA is not:
- Not a pupil-facing chatbot or tutor.
- Not a free-text conversation between a pupil and a model.
- Not a tool that collects pupil personal data to function.
- Not an engagement product — no streaks, nudges or retention loops.
- Not a replacement for a school’s filtering, monitoring or safeguarding.
5. Status note for readers
EVA is in pre-launch. Items marked in progress reflect documents being finalised ahead of launch, not gaps in the product’s design. Each will carry a published link and date as it is completed. The Data Protection Impact Assessment and the question of whether any feature brings EVA within scope of the Online Safety Act 2023 are being confirmed with a data-protection adviser before this mapping is treated as final.
6. Contact and maintenance
Grey Gecko Educational Technology Ltd. Compliance enquiries: faye@greygecko.ai. Accountable owner: Faye Fulton, Founder.
This mapping is reviewed against the gov.uk source each cycle (next review: 24 December 2026). Standards version referenced: 19 January 2026. The DfE revises this collection roughly quarterly — re-verify before quoting any individual line.
This page is a good-faith conformance mapping prepared by EVA™ to assist schools and colleges in their own assessment. It is a summary, not legal advice, and does not by itself constitute approval of EVA for use in any setting. Schools remain responsible for their own due diligence, data-protection decisions and safeguarding duties under Keeping Children Safe in Education. The authoritative statement of the standards is the Department for Education publication.