"

[!IMPORTANT] Defining AI Safety Safety, in the context of the Oxford Rubric, means protecting pupils from physical, psychological, emotional, and informational harm when AI is used in any school setting. It includes data protection compliance, age-appropriate content moderation, and clear protocols for when things go wrong — and it cannot be delegated to the technology itself.

Of the five criteria that make up the Oxford Rubric — the principled framework for assessing AI use in schools — Safety is the one most schools assume they've already covered. A safeguarding policy exists. GDPR training has been done. The IT team checked the terms and conditions.

But the Oxford Rubric sets a more demanding standard than most schools realise. And when you look at what Safety actually requires, the gap between "we've thought about it" and "we've genuinely addressed it" becomes clear.

What Does the Oxford Rubric Mean by Safety?

The Oxford Rubric defines Safety as the extent to which pupils are protected from physical, psychological, emotional, and informational harm when AI is implemented in any school scenario. It also covers how safeguarding duties are upheld and data protection obligations are fulfilled.

Three things in that definition are worth pausing on. First, it covers all four types of harm — not just the informational risks (inappropriate content) that usually dominate the conversation, but psychological and emotional risks too. An AI tutoring tool that responds to a student's distress without flagging it to a human, or that creates unhealthy dependency over time, is a Safety concern even if it never surfaces harmful content.

Second, it includes foreseeable misuse and unintended consequences — not just intended functionality. A well-designed AI writing tool can still become a route to harm if students use it to generate content that undermines their own thinking, or if it's used in a way that the school hasn't anticipated.

"

"The rubric is explicit that safeguarding responsibilities cannot be delegated to technology. However sophisticated the AI, the school's non-delegable duty of care remains with the adults in the institution."

What Does a Safe Use of AI Actually Look Like?

The Oxford Rubric identifies several clear indicators of appropriate practice:

"

[!TIP] Actionable Safety Measures Ensure a documented child-centred safeguarding and data protection impact assessment has been completed for your specific AI use case, age group, and data involved. Generic assessments aren't enough — the assessment must reflect the actual tool, the students using it, and the context in which it's being used.

• A documented child-centred safeguarding and data protection impact assessment has been completed. Not a generic DPIA — one that specifically considers the use case, the age group, and the data involved.
• Age-appropriate content filtering, moderation, and guardrails are enabled by default. Not as an optional add-on, and not relying on students to self-regulate.
• Human supervision is maintained for all high-risk or sensitive interactions. AI should not be the only "person" a student interacts with in any situation involving wellbeing, pastoral care, or sensitive topics.
• Data collection is minimal, proportionate, and clearly justified educationally. If a tool collects more data than the educational purpose requires, that's a red flag.
• Clear procedures exist for reporting, responding to, and learning from incidents.

This is precisely why platforms like Secondary AI are designed with teacher oversight at their core. Our chatbots support AI-guided inquiry and provide immediate feedback to students — but any flagged content or signs of disengagement are surfaced directly to the teacher for human review. The AI never operates as the sole responder in a sensitive interaction. Human supervision is always maintained.

Equally instructive is what the rubric flags as inappropriate: pupils interacting with AI without adequate moderation or supervision; sensitive pupil data processed opaquely; safeguarding responsibilities implicitly shifted onto the system; and failure modes that haven't been considered in advance.

The Questions Secondary Schools Should Be Asking

For school leaders and teachers evaluating any AI tool or practice, the Safety criterion translates into a set of concrete questions:

• Does this AI system comply with child-specific data protection standards — not just general GDPR?
• Are the content filters enabled by default, and are they calibrated for our students' age group?
• If a student discloses something concerning to this AI, what happens? Is there a human in the loop?
• Could reliance on this system create emotional harm, dependency, or a sense of exclusion for any students?

"

[!WARNING] The Unanswered Question Most AI procurement conversations focus on what the tool does when it works. But schools must also ask: What happens when the AI fails, produces an error, or is deliberately misused? Do we have a protocol? Failure to answer this before deployment isn't an oversight — it's a safeguarding gap.

Why Safety Is the Non-Delegable Starting Point

The Oxford Rubric makes clear that all five criteria must be met for AI use to be considered appropriate — failure in any one signals a need for redesign, restriction, or rejection. But Safety has a particular weight because of the statutory context of schools.

Secondary schools have legal duties around safeguarding, data protection, and duty of care that adult workplaces do not.

"

"The children in your care are at a stage of development where they are more vulnerable to psychological harm, more susceptible to unhealthy dependency, and less equipped to critically evaluate what they're being told — by a person or an AI."

This is why the rubric frames Safety as a non-delegable responsibility. A vendor can tell you their product is safe. But the school is the one that needs to have verified that claim, documented the assessment, and put human oversight in place before any student interacts with the tool.

Secondary AI adheres strictly to this principle. Our platform is designed to empower teachers with AI support, not replace their essential role. Safety protocols are always teacher-controlled — from content moderation settings to the review of flagged conversations — ensuring that human oversight is never optional, and never assumed.

The other four criteria of the Oxford Rubric — Efficacy, Accountability, Transparency, and Agency — build on Safety as the foundation. But Safety is where every evaluation has to start.

Frequently Asked Questions

What is the Safety criterion of the Oxford Rubric?

"

[!QUOTE] Safety Criterion — Oxford Rubric The Safety criterion assesses whether pupils are protected from physical, psychological, emotional, and informational harm when AI is used in school. It includes data protection compliance, appropriate content moderation, human supervision for sensitive interactions, and clear incident protocols. It cannot be satisfied by vendor assurances alone — schools must complete their own child-centred assessment.

How is AI safety in schools different from general data protection compliance?

General data protection compliance addresses how personal data is collected, stored, and processed. AI safety in a school context goes further, covering the psychological and emotional impact of AI interactions, the risk of harm from AI outputs, the appropriateness of content for students' age and developmental stage, and what happens when the AI fails or is misused. The Oxford Rubric requires all of these to be addressed, not just data handling.

Can an AI tool be used in school if it doesn't fully meet the Safety criterion?

No. The Oxford Rubric states that a use case should only be considered appropriate if it satisfies all five criteria. An amber rating on Safety — meaning risks are partially mitigated but not eliminated — requires additional work before the tool is deployed with students. A red rating means the use is currently inappropriate.