An AI Policy Teachers and Students Can Actually Follow

A teacher uses a tool to suggest examples for a lesson. A learner asks it to explain a difficult passage. An administrator uses it to draft a message about a student's circumstances. These activities involve different information, different risks and different educational purposes.

Treating them as one category makes a policy difficult to follow. Staff may assume that an approved product permits every use. Learners may interpret permission to get help as permission to submit a generated answer. Managers may expect teachers to resolve the ambiguity case by case.

Begin with the work being proposed. State its purpose, the person responsible and the result expected. Then decide which parts the tool may assist and which parts require a human judgement or an independent learner response.

UNESCO's guidance on generative AI in education emphasises privacy, human agency and age-appropriate educational design. An institution needs to translate those concerns into instructions people can use during an ordinary lesson or administrative task.

The policy should answer practical questions. Can a teacher upload student work? Can a learner use generated feedback before resubmitting? Can a tool propose a grade? Who checks a draft communication before it reaches a family?

Clarity does not require a rule for every imaginable prompt. It requires a small set of well-defined uses, an explicit boundary around sensitive work and a route for reviewing proposals that fall outside the current rules.

A workable starting policy can separate four activities. The conditions below are an editorial model to adapt with the institution's teaching, data and safeguarding leads.

The product's presence on an approved list is only one condition. An approved account may still be unsuitable for a particular document, learner age or assessment purpose. Conversely, a cautious policy can permit a narrow use without endorsing every feature of the service.

Write the boundary next to the activity. A teacher preparing materials should see the relevant data and checking rules without searching a long policy document. A learner should know the permitted help before starting the assignment.

Include a route for exceptions. A proposed new use should identify what changes, why it helps and what additional review it needs. This makes the policy capable of learning while keeping responsibility with the institution.

The person reviewing an output needs enough subject knowledge and context to judge it. A requirement to “check the AI” is weak if the reviewer cannot identify a plausible error or does not have time to examine the material.

For lesson preparation, decide what checking involves. The educator may need to verify factual claims, work through answers, inspect examples for misleading assumptions and confirm that the language suits the group. A generated reference should lead to a source that actually exists and supports the point.

NIST's generative AI risk profile identifies confabulation and privacy risks. In an education workflow, those concerns need a concrete response rather than a general warning placed at login.

Consider a hypothetical tutoring provider preparing reading exercises. A tool produces a fluent passage and comprehension questions. The tutor checks the passage, discovers that one answer relies on information never supplied and revises the exercise before learners see it. The value of assistance depends partly on that review effort.

Make it possible to record and correct recurring problems. If a tool repeatedly produces unsuitable reading levels or unreliable explanations in a subject, the institution should narrow the use or stop it while the issue is investigated.

Give learners a way to question automated assistance. They should know where to report an answer that conflicts with taught material or feedback they cannot understand. An educational service must remain open to correction, including when its output appears confident.

A student's work can contain much more than an answer to a question. It may reveal a name, personal circumstances, learning needs or information about someone else. Removing the name alone may leave the person recognisable.

Specify which information may enter each service and for what purpose. The review should cover account arrangements, retention, access, any use of submitted material for model improvement and the institution's ability to retrieve or delete records. Staff need a plain answer for the tool and task they are using.

Use less information where the task permits it. A teacher seeking alternative examples may be able to describe the learning objective without uploading individual student records. An administrator drafting a routine notice can work from a neutral template before adding personal details within an approved system.

Avoid making learners discover the rules through trial and error. Explain which accounts they may use, what they should keep out of a prompt and where they can obtain help without sharing personal information.

Age, consent and parental involvement require local and institutional review. A service's minimum age, an educational recommendation and a legal requirement answer different questions. Permission under one does not automatically settle the others.

Keep records proportionate. Oversight does not require copying every learner conversation into a permanent staff archive. Decide which evidence is needed to investigate incidents or assess a pilot, who can access it and when it should be removed. The monitoring arrangement deserves the same care as the original use.

An assessment policy should start with what the task is intended to establish. If it tests a learner's ability to construct an argument, unrestricted generation of the argument changes the evidence. If it tests critical evaluation, reviewing an AI-produced answer may be part of the task itself.

State the permitted assistance in the assignment. Distinguish help with understanding instructions, generating ideas, editing language and producing the substantive answer. Learners need examples that make those distinctions intelligible.

Where assistance is allowed, decide what acknowledgement is appropriate. The purpose is to understand how the work was produced and whether it demonstrates the intended ability. A disclosure rule should be clear enough to follow without turning every minor interaction into an administrative exercise.

Provide opportunities to show understanding through the work. An educator might ask for an explanation of choices, a revision in response to feedback or an application to a changed problem. The approach should fit the subject and the learners.

Keep concerns about misconduct within a fair institutional process. A learner should be able to explain their work and challenge an adverse judgement. The institution needs a responsible person who evaluates the available evidence.

The EEF's EdTech research agenda explicitly examines which thinking processes AI may support and which learners need to retain. That question belongs at the centre of assessment design. The educational purpose should determine the boundary around assistance.

A first deployment needs a defined purpose and a review date. Select a limited use, prepare the people involved and establish how they will report errors or uncertainty. Make a supported alternative available when the tool cannot be used.

Review the work produced, the checking effort and the experience of the learners or staff involved. If the intended benefit is reduced preparation time, include review and correction in the comparison. If it concerns learning, examine evidence of understanding separately.

Assign someone to maintain the policy and the approved-use list. Product features, account terms and institutional requirements can change. A previous approval should have a review path when those conditions shift.

Publish revisions in language that explains the practical change. “You may now use this tool for this activity, with these conditions” is easier to apply than a broad statement about innovation.

The institution should be able to account for where AI contributed, who remained responsible and how problems were corrected. That is a firmer basis for wider adoption than a policy that offers permission without explaining the work.