Governing AI on Paper and in Practice
Teachers are using AI faster than schools are writing rules for it. This paper is about closing that gap well: not with a longer document, but with a policy that makes a small number of real decisions, lives inside the tools teachers use, is taught to the people it governs, and is looked at again every term. Like the papers before it, it names no vendor as an adversary.
النسخة العربية من هذه الورقة ←
Read the full paper
Tell us who you are and the rest of this paper opens straight away. It is free.
AI arrived in classrooms before the rules did.
The evidence points one way. RAND's national survey panels in the United States found that in 2025, 54 percent of students and 53 percent of English language arts, mathematics and science teachers used AI for school, increases of more than 15 percentage points on surveys one to two years earlier. Yet only 45 percent of principals reported school or district policies or guidance on the use of AI, and 34 percent of teachers reported policies on AI and academic integrity. The authors' title states the finding plainly: AI use in schools is quickly increasing, but guidance lags behind (Doss et al., 2025).
UNESCO saw the same gap from the other side. Its global guidance observes that new versions of these tools are released faster than rules can follow, and that without national regulation educational institutions are "largely unprepared to validate the tools" (UNESCO, 2023).
The Gulf is not waiting. The UAE Ministry of Education has introduced artificial intelligence as a formal subject in public schools from kindergarten to Grade 12, beginning gradually in the 2025–26 year, with ethics making up a quarter of the curriculum and around 1,000 trained teachers delivering it (Rasheed, 2025). When pupils study AI as a subject, a school can no longer treat its own use of AI as somebody else's question.
UNESCO is clear on where the responsibility sits. Governments regulate; and the guidance "stresses the need for educational institutions to validate GenAI systems on their ethical and pedagogical appropriateness for education" (UNESCO, 2023). A school AI policy is where that responsibility becomes concrete.
A policy that lives in a PDF governs nothing.
Schools that do have an AI policy often discover that it changes very little. Four failures recur.
- Written for lawyers, not teachers. A policy that reads like a contract is filed, not used. The teacher deciding at 9 pm whether to paste a pupil's notes into a chatbot needs one sentence, not ten pages.
- Not enforced where the work happens. If the rule is "only approved tools" but nothing stops an unapproved one, the rule is a hope. Perspectives No. 3 called the result shadow use: AI the school cannot see.
- Sent, not taught. A policy emailed to staff is read by a few and argued about by none. RAND's data make the point with numbers: among teachers who had access to AI training, 37 percent found their AI policies helpful; among teachers without it, only 13 percent did, and 53 percent found their policies unhelpful. In the authors' words, "AI policies are likely more helpful when they are combined with training" (Doss et al., 2025).
- Never reviewed. AI tools and the models behind them change within a school year. UNESCO asks that a tool in use be "carefully monitored and validated" throughout its life cycle, not approved once (UNESCO, 2023), and Perspectives No. 4 showed that the AI a school approved in September may not be the AI drafting its lessons in December.
A useful policy is short because it decides things.
| Decision | What the policy should say | How to make it real |
|---|---|---|
| 1. Which tools | The AI tools staff may use for school work, and that others are not used with school data | A short approved list, reviewed termly; a governed channel teachers prefer |
| 2. Which data never goes in | A plain list: full names with grades, health, family circumstances, ID numbers, anything a parent would not expect to leave the school | Tools that remove or mask names before any AI call, and a "never enter" list on one page |
| 3. Who reviews AI output | Nothing AI-drafted reaches a pupil or parent until a teacher has read and approved it | An approval step in the tool, recorded with who and when |
| 4. What students may do | When pupils may use AI, under whose supervision, and how they say they used it | Classroom rules on the wall, taught in a ten-minute lesson |
| 5. What parents are told | What the school uses AI for, what it never does, how a child's data is protected, and whom to ask | A one-page letter at the start of term, in Arabic and English |
Each row is a decision a head can take in a meeting. None needs legal drafting to begin; each needs the school's own advice on the law that applies to it.
The strongest policy is one the tool already obeys.
Decisions 1 to 3 are far easier to keep when the tool teachers use carries them out. A school choosing any AI platform should ask whether it can:
- set the boundaries once for every teacher: which languages, which question types, how long an assessment may be, what content guidance applies, and which AI providers may receive the school's requests;
- keep student names away from the AI: replacing them before a request is sent and restoring them only in the teacher's draft;
- require a teacher's approval before a draft is printed, downloaded or exported, and record it.
- check what the AI claims, before a teacher sees it: a model that says a lesson meets an official objective should not be the only one who checked.
In EdSmartly, the school sets thirteen governance controls once; six of them are enforced by the server on every generation (assessment length, a per-teacher daily limit, question types, output languages, content guidance and AI providers). Student names are replaced with neutral placeholders before any request leaves the platform, and restored only in the draft the teacher sees. A school can require approval before printing, download or export. That is a safeguard, not a lock: copying stays possible, because the teacher is a professional.
One rule in EdSmartly used to rest on the model's good behaviour, and we changed it in October. Every draft ends by listing the curriculum objectives it covers, each marked Verified (taken from the school's own curriculum) or Indicative (the AI's suggestion). The AI was told to mark an objective Verified only when it came from the school's list. Being told is not the same as being checked. Now the server checks every Verified label against the codes it gave the AI, and any it cannot match becomes Indicative, with the reason written beside it, before the teacher sees the draft. The school's dashboard and its governance sheet show how many labels were checked and how many were corrected in the last thirty days. A school does not have to trust that the AI followed the rule; it can see the count.
The point is not this product. It is that a rule a tool enforces is a rule a school can stand behind.
Staff, pupils and parents each need the policy in their own form.
- Staff: a ninety-minute session beats a long email. Show the tool doing something useful, show it being confidently wrong, put the "never enter" list on the table, and leave with the decisions written down.
- Pupils: six lines on the classroom wall, and a short lesson that makes them mean something, including the moment a model gets a fact wrong with complete confidence.
- Parents: one letter, at the start of term, that says plainly what AI is used for and what it is never used for.
UNESCO puts the same point in institutional terms: schools need to understand the benefits and risks of AI before they can validate its adoption, and teachers "need to be supported to strengthen their capacities for the proper use of GenAI, including through training and continuous coaching" (UNESCO, 2023). The aim is a staff room that has argued about the policy, because a policy people have argued about is a policy people remember.
Write it in both languages
In a Gulf school the policy should exist in Arabic and in English from the first draft, not translated at the end. Staff, pupils and parents read it in the language they think in, and a rule that is clear in one language and vague in the other will be followed in only one. Take advice on which text prevails if the two ever differ, and keep them in step whenever the policy changes.
AI changes during the school year; the policy should keep up.
Put a twenty-minute review on the calendar each term. Three questions are enough:
- What changed in the tools? Ask your provider for a dated record of changes to the models behind your teachers' work.
- What did teachers and the checks find? What teachers edited, what they stopped using, what they wished they could do; and, where the tool counts it, how often the AI's claims had to be corrected. A rising count is a reason to ask the provider what changed.
- What did parents and pupils ask? Every question is a sentence the next letter should answer.
Record the answers and the date. A policy with a review history is a policy a school can show it keeps.
Five moves for heads and ministries
- Decide the five things, this term. Tools, data, review, pupils, parents. One page.
- Choose tools that enforce the rules. Prefer platforms where the school's settings apply to every teacher, names are masked before AI, approval is recorded, and the AI's claims about the curriculum are checked by the platform, not taken on trust.
- Train before you send. Run one staff session before the policy goes out, not after.
- Tell parents first. A letter at the start of term earns more trust than an answer after a complaint.
- Review every term. Ask what changed in the tools, in teachers' use, and in parents' questions, and write it down.
The AI proposes; the teacher decides
EdSmartly's position has not changed since No. 1: the AI proposes; the teacher decides. A policy is how a school writes that down for its own staff, pupils and parents.
To help a school start, we publish the School AI Policy & Staff Training Kit: an acceptable-use policy, one-page staff guidance, a parent letter, classroom rules with a ten-minute lesson, and a ninety-minute staff session outline, in English and Arabic, editable in Word. It is templates to adapt, not legal advice; the school stays responsible for the policy it adopts. A school that wants to see the rules enforced in practice can run a one-term pilot with up to twenty teachers.
An invitation
If the question in this paper is one your school or authority is asking, we would welcome the conversation.
References
- Doss, C.J., Bozick, R., Schwartz, H.L., Chu, L., Rainey, L.R., Woo, A., Reich, J. and Dukes, J. (2025) AI use in schools is quickly increasing but guidance lags behind: findings from the RAND survey panels. RR-A4180-1. Santa Monica, CA: RAND Corporation. Available at: www.rand.org/t/RRA4180-1.
- Gallot, H. (2026) 'Securing the agentic insider', Official Microsoft Blog, 19 October. Available at: blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/.
- Rasheed, A. (2025) 'UAE schools to introduce AI curriculum in new academic year', Gulf News, 23 August.
- UNESCO (2023) Guidance for generative AI in education and research. Drafting led by F. Miao, co-drafted by W. Holmes. Paris: UNESCO. https://doi.org/10.54675/EWZM9535.
Contact
hello@edsmartly.com · EdSmartly.com · app.edsmartly.com
EdSmartly is a service of Axiom Intelligence Partners LLC.
This document may be shared freely in unmodified form.