Using Gemini safely in schools: four rules that work better than a ban

Picture a colleague who tells you they have been using Gemini to plan their lessons. They generate the plan, glance over it, and teach it. How do you feel about that?

We open our staff AI training with that question, most recently at an INSET session for a Hull academy trust, and the room always splits. Some people feel the colleague is cheating. Some feel they are being efficient. Almost nobody can say precisely where the line is. That is the real AI problem in schools right now: not the technology, but the absence of a shared professional standard for using it. This article sets out the one we teach.

Why banning AI in schools fails

The instinct in some schools is still to prohibit. It fails for a simple reason: staff are using AI anyway, because it saves them hours, and a ban does not remove the use, it removes the visibility. The worst possible arrangement is the one a ban creates: staff pasting school information into free personal chatbots on their own devices, with no data protections, no oversight and no shared standard, while everyone pretends it is not happening.

The workable position is the opposite one. Give staff a sanctioned tool inside the school's managed environment, and pair it with rules that are short enough to remember and reasonable enough to follow. This is not permissiveness. It is the only version of control that actually exists.

What Google's education environment actually protects

For schools on Google Workspace for Education, the sanctioned tool question has a clear answer, because Gemini and NotebookLM are core services under the education terms of service. In plain English, that means prompts and uploads are not reviewed by humans and are not used to train Google's public models, and interactions stay within the school's own environment. Editions like Education Plus extend this from the standalone Gemini app into the tools staff already live in, with Gemini working inside Docs, Gmail and Slides.

That is a categorically different arrangement from a free consumer chatbot on a personal account, and it is worth being precise about why: the difference is not the model, it is the terms. Same underlying AI, completely different data relationship. When we describe a school's Education environment as a walled garden, that is the claim: your data stays yours.

One important nuance for leaders: this protects data, not judgement. A walled garden stops your information leaking; it does not stop a member of staff putting a child's name and diagnosis into a prompt, and it does not make the output accurate or appropriate for your pupils. The environment is the foundation. The rules are the building.

The four risks the rules exist to manage

When we train staff, we start with risks rather than rules, because a rule you understand is a rule you keep. There are four, and none of them will surprise anyone who has used these tools.

Deskilling. If AI plans every lesson for two years, who still knows the curriculum progression cold? The risk is not that AI does the work badly. It is that it does the work well enough that professional knowledge quietly atrophies.

The average trap. AI produces the average of everything it has seen. It is competent and generic. It does not know Hull, your community, or which class is hard work on a Friday afternoon, and unedited output reads exactly like what it is.

Hallucination. Generative AI predicts plausible text; it does not verify truth. It is fluent, confident and sometimes flatly wrong, which matters more in front of thirty children than almost anywhere else.

Data privacy and safeguarding. Even inside a protected environment, typing a pupil's name, diagnosis or home situation into a prompt is bad practice, because it builds habits that will eventually be exercised somewhere less protected.

The four rules of engagement

The rules answer the risks one for one, and they are deliberately short.

1. The 80/20 rule. AI drafts the 80 per cent; the member of staff supplies the 20: reviewing, editing and fact-checking against the school's own curriculum standards. Nothing AI-generated goes in front of pupils or parents without active human review. The point is not that the output is bad. The point is that the edit is where professional expertise lives, and it is not optional.

2. Anonymise first. No personally identifiable information in prompts, ever, regardless of how protected the environment is. Not "write a report for Sarah Smith, who struggles with dyslexia and keeps disrupting maths" but "a Year 4 pupil overcoming reading difficulties who needs to focus on engagement". Same usefulness, zero exposure, better habits.

3. Radical transparency. If a teacher uses Gemini to draft a model essay, they tell the class, and they critique it together. "I generated this with AI, let's find its weaknesses" is not an admission, it is a lesson in exactly the critical thinking pupils will need. Hiding AI use teaches the opposite.

4. Contextual anchoring. Do not ask for "a six-week unit on the Romans". Paste your actual objectives and ask for materials based only on those. Generic requests produce the average trap; anchored requests produce something usable. This rule is also the answer to the colleague in the opening question: a fully AI-generated plan is not banned, it is unfinished until it has the 20 per cent edit, no PII, honest attribution and curriculum grounding.

Our Digital Transformation Lead, Devin Falvey, who delivers these sessions, describes the pattern he sees: "Schools are eager to dive into AI without fully understanding how rapidly it is evolving. Our job is to make sure schools are not only using AI safely, but using it in a way that has a lasting impact." That second half is the point of the four rules: they are not restrictions on AI use, they are what makes the use worth anything a year later.

Two tools, and which to reach for

One refinement worth teaching explicitly: Gemini and NotebookLM are different tools for different jobs. Gemini is the broad assistant, drawing on the open web, best for differentiated texts, quiz questions and communications, and everything factual it produces gets verified before use. NotebookLM is the closed-book synthesiser: it answers only from documents you upload, which makes it the right tool for curriculum-locked materials and specification summaries, with far less hallucination risk but limited to your sources. Broad and creative, Gemini. Locked to your curriculum and accurate, NotebookLM. Staff who know which to reach for stop fighting the tools.

Making it stick: policy, training and the leadership piece

A framework only becomes a professional standard when three things line up. The policy work: a short, principle-based acceptable use position, owned at leadership level, consistent with the DfE's position on generative AI and folded into the school's existing digital policies rather than bolted on. The technical work: access decisions made deliberately in the admin console, because Gemini access for schools is an explicit choice, with age-based restrictions applying to under-18 users, and pupil-facing use is a separate decision from staff use with its own safeguarding considerations. And the human work, which is the one that actually changes behaviour: proper staff training, delivered as CPD with hands-on practice, not a policy PDF attached to an email.

This is the digital literacy pillar of our five-pillar approach for schools and trusts, and it is the pillar most AI conversations skip: the technology decision is easy, and the people decision is the whole game. We covered how the pillars fit together in our article on meeting the DfE digital and technology standards.

Where to start

If your school or trust has Google Workspace for Education, you already own the safe environment; what most settings are missing is the professional standard and the confidence. We deliver the session this article is based on as staff INSET, hands-on with Gemini and NotebookLM, built around the four rules and adapted to your setting. Get in touch to talk about running it for your staff.

Frequently asked questions

Is Gemini safe for schools to use?

Within Google Workspace for Education, Gemini operates as a core service under the education terms: prompts and uploads are not human reviewed and are not used to train Google's public models. That makes the environment safe; safe use still depends on staff practice, which is what an acceptable use standard and training provide.

Does Gemini use school data to train Google's AI?

No. For Google Workspace for Education customers, Gemini interactions stay within the organisation's environment and are not used to train models outside it. This is the key difference between a school-managed Gemini account and a free personal chatbot account, where different terms apply.

Can pupils use Gemini in school?

Access is controlled by the school's administrators and is off by default for new primary and secondary customers, so pupil access is a deliberate decision, not an accident. Certain features are restricted for users under 18, and pupil-facing use carries its own safeguarding and curriculum considerations beyond staff use.

What should a school AI policy include?

Keep it short and principle-based: which tools are sanctioned, what information may never go into a prompt, the expectation of human review before anything reaches pupils or parents, and transparency about AI use. Four memorable rules that staff actually follow beat ten pages nobody reads.