Your AI policy won't save you, capacity will: Why compliance is the wrong goal for schools.

A few numbers frame the moment we are in. The AI market in K-12 education is growing at a 44% CAGR, on its way to $9.7B by 2030. Nearly two-thirds of students now use AI for their schoolwork, and 60% of teachers use it in theirs, almost double two years ago. And yet fewer than 10% of schools have any formal AI policy in place.

Read them together and the picture is hard to miss: adoption has outrun governance. Students, teachers, and increasingly leaders themselves are all using these tools daily, just without a shared policy to lean on. The question is no longer whether to write one, it's how to write one people will actually use.

And the problem isn't unique to AI. Pick any policy your team worked hard on: assessment, language, teaching and learning, and ask a teacher what it says. You'll usually get a shrug: they know it exists but rarely reach for it, because digging it out mid-lesson costs time they don't have. So they fall back on something simpler, common sense and whatever feels relevant in the moment. That's worth sitting with before writing a word about AI: the best preparation isn't drafting, it's reflecting honestly on how your existing policies are doing. The divide that matters isn't whether you have a policy, it's whether it gets used.

It helps, then, to be clear what a policy really is. The document is only the surface; the real policy is the set of shared expectations people act on, carried less by publication than by the routines, training, and everyday conversations around it. At root, it's your school's mission and values made specific for a situation. When it's working, a teacher facing a hard AI call doesn't hunt for the document, they already sense what the school would want.

So here are my two cents for building a policy your whole school will actually adopt, less about predicting where the technology goes, more about getting the foundations right.

Get the document right

We don't have to guess at how an AI policy might go wrong, we've already watched it happen with the policies we have. The same handful of mistakes recur. Here are 4 principles for avoiding them, each the flip side of a failure I've seen again and again:

1. Build it together. Policies drafted by a small group and handed down tend to answer questions no one is asking while staying silent on the ones that come up daily. So bring teachers and support staff into the drafting, not just the rollout, and listen just as carefully to students and parents, who are central voices, not a courtesy consultation. Students see how AI actually shows up in their work; parents bring expectations the school can't afford to guess at. Remember, too, that leadership lives by the same policy, not above it. When the whole community shapes it, the policy earns the shared ownership that makes people reach for it rather than set it aside.
2. Be specific. "Use AI responsibly." "Apply sound professional judgment." Phrases like these feel safe because they commit to nothing, but they give a teacher who needs a quick decision nowhere to stand. Pair every principle with a concrete example: what good use looks like, what crosses the line, and a clear answer for the situations that come up most. That kind of specificity is only possible once you genuinely understand the tools in play, what the school has adopted, what staff quietly use to plan and mark, what students turn to. Catching yourself reaching for a vague phrase is itself a signal: it usually means you don't yet understand that area well enough to be specific, the cue to go and learn before you write.
3. Offer a path, not just a ban. Blocking tools rarely stops people using AI, it just drives the use into the dark, with no guidance and no shared norms - worse than no policy at all. A restriction can be the right call, but only with the reasoning behind it and a sanctioned route to whatever people were reaching for: a maintained list of approved tools, and a simple, visible way to request new ones with a turnaround a busy teacher can plan around. Set clear criteria, and treat a data-protection agreement as non-negotiable for anything that touches student information, more layered still in an international school juggling GDPR and other regimes at once. Being constructive beats enforcing: a path people can follow does far more than a rule you have to police.
4. Fit the context. AI in an early years literacy block is a different animal from AI in an IB Extended Essay. Take a blanket "no AI on written work": for a Year 12 student it bans the supported drafting they'll be expected to do at university; for an EAL student it treats a legitimate language scaffold as cheating; for a six-year-old it's beside the point. One sentence, three needs, none of them met. Add a multilingual student body and a parent community with varied expectations, and a uniform rule fits almost no one. The fix is inclusive drafting, staff from every phase and department in the room, so the policy is built with each context rather than guessing at most of them.

None of these are new. They're the patterns that have undermined good policies for years. Get them right and you'll have a document worth adopting. But the document is only half the job. The other half is making it live.

Make it live

A document is only as good as people's ability and willingness to use it. Two things turn a written policy into a living one.

1. Build the capacity to use it. A policy that asks teachers to "exercise professional judgment" about AI is asking for a skill most were never trained in, and that judgment doesn't come bundled with the PDF. Without it, even a well-written policy stays inert. So treat professional development as part of the policy, not a footnote: subject-specific rather than generic, low-stakes enough that people can experiment and ask honest questions, and - crucially - continuous, since a single session in August teaches tools that will have changed by half-term. It works best when leaders learn alongside staff rather than sending them off to it, guidance from someone who has never used the tools drifts from how they actually behave.
2. Keep it current. Most school policies can sit for years between reviews; an AI policy can't. The tools change month to month, so anything you publish today will need revising within the year. This is the one policy to treat as a living draft. Build that in and say so plainly: review it once a term (or even more frequently) rather than once a year, give it one named owner rather than a committee, and keep a feedback channel teachers can reach in seconds. A policy that pretends to be final loses credibility the moment reality moves past it; one that visibly evolves earns the trust that gets it used.

The hardest case: academic integrity

All of this gets hardest in one place, and it deserves a direct answer rather than a hedge: academic integrity, the issue that drives most of the anxiety in these conversations. The honest starting point is that no one has this fully worked out: the whole sector is adapting in real time, and that uncertainty is something to tolerate and learn within, not to paper over with false certainty. Hold the principle firm and stay flexible about the approach. The principle is academic integrity. How you manage it should be a fair, human process, not just a detection tool.

Yes, students are using AI to do work they were meant to do themselves, and yes, that is a real problem worth taking seriously. But two things are just as true, and both belong in your policy. First, AI detection tools are unreliable, and they are especially prone to false positives on the writing of non-native English speakers: a widely cited Stanford study found detectors wrongly flagged more than half of essays by non-native writers as AI-generated, while almost never misjudging native speakers. In an international school, where a large share of your students are writing in their second or third language, that is not a hypothetical risk, it is a near-certainty of real harm. So don't let a detector's score be the verdict. Your policy should say plainly that detection output alone is not evidence, and it should set out the procedure that is: a conversation with the student, a look at drafts and process, and professional judgment applied consistently, a fair process every student can count on, rather than a number nobody can defend.

Second, the durable answer is not surveillance. It is rethinking how we assess. Part of that is making individual tasks harder to outsource, work built around in-class writing, oral explanation, a student's own context and experience, projects that develop visibly over time, and reflection on thinking rather than just a final answer. But the bigger shift is asking what's worth assessing at all in a world where AI can produce a polished essay on demand. If a tool can do the task, the task may no longer be measuring what matters. That pushes us toward assessing the things AI can't fake: a student's reasoning, their judgment, how they question and improve on what a machine gives them, the process rather than just the product. Some schools are even beginning to assess how well students use AI itself, since that is fast becoming a real-world skill. A useful test for any existing assignment: could a student get a strong result by pasting it into a chatbot, and if so, what would make that impossible? Adding an oral defence, asking for drafts and process alongside the final piece, or grounding the task in something only that student has - their own data, a class discussion, a local context - all help. Helping teachers modernise assessment this way is a much better investment than any enforcement system. The goal isn't to catch students afterward. It is to design learning worth doing in the first place.

The leader's job: learn alongside

The leaders getting this right are not the ones who shipped the longest, most detailed policy the fastest. They are the ones who created room for an honest professional conversation. In effect they told their faculty:

We are working this out together. Here are the principles we are anchoring to. Here is what we know and what we are still figuring out. And here is how we will support you while it changes.

That posture is not weakness. It is accuracy, and your teachers can tell the difference between a leader genuinely learning alongside them and one performing a confidence they don't actually have. The schools that bring teachers in early, tie policy to real development, keep it breathing, and treat teacher judgment as a skill worth growing will end up with far more than a document.

They will have teachers who are genuinely ready to prepare students for a world where AI is already everywhere. That is the goal. Not compliance. Capacity.