Teaching · Practical guide
Draft Practical Institutional AI Guidelines
Turn broad principles into usable rules for data, assessment, procurement and responsibility, with an approval-ready example and scenario checks.
Define scope and authority first
A guideline needs to say who it applies to, which activities it covers and who can approve exceptions. A teacher’s classroom note is not automatically an institution-wide policy. This lesson helps you prepare a proposal for the appropriate review process; it does not create binding rules for any institution.
Separate staff preparation, student learning, assessed work and administrative decisions. These uses involve different data and consequences. A tool that is acceptable for public lesson examples may not be approved for student records or automated decisions. State the scope clearly enough that a reader can identify which rule applies to a concrete task.
Translate principles into decisions
“Protect privacy” is a principle. “Use synthetic examples in unapproved tools; obtain the designated approval before processing restricted institutional data” is an actionable rule. “Maintain human oversight” becomes more useful when it names who reviews an output and which actions cannot occur automatically.
A draft should cover allowed data, approved services, access arrangements, output verification, assessment disclosure, incident reporting and periodic review. UNESCO’s education guidance and teacher competency framework offer relevant background. They do not replace your institution’s legal, security, accessibility or academic approval processes. Record the local decisions that still need an authorised owner.
Worked example: a teacher preparation workflow
Proposed scope: drafting practice materials from public course concepts. Permitted input: public definitions and fictional examples. Restricted input: names, IDs, grades, personal circumstances, passwords and unpublished assessment banks unless a specifically approved arrangement permits the use. Required check: the teacher verifies calculations, references and alignment before release.
Proposed output rule: AI can draft questions and feedback patterns; it does not approve an assessment or send student-specific decisions. Owner: the course team reviews educational suitability; designated institutional teams approve service and data conditions. Record date, version and review trigger so the rule can be revisited when the tool or its use changes.
Scenario test: a teacher wants to upload a named grade spreadsheet to draft comments. This falls outside the public/synthetic preparation rule. The draft should direct the teacher to the approval process or a fictional practice version, not assume removing the name column makes every remaining record safe.
Plan for access, exceptions and incidents
A policy requiring every student to use a paid product can create an access barrier. Define a viable alternative and check whether the intended activity can be completed without personal payment or unnecessary account creation. Consider accessibility and age-related arrangements through the institution’s established processes rather than asking a model to determine eligibility.
State where to report accidental disclosure or a harmful output. The response should identify the service, time and information category without spreading the sensitive material further. The relevant team decides containment and notification steps. Do not promise that deleting a chat reverses a disclosure or that a setting guarantees all data-handling requirements are met.
Test the draft with cases and assign owners
Use short scenarios: public lesson brainstorming, student work containing personal data, generated final assessment answers, a tool requesting broad access and an inaccurate citation. For each, ask a reader to identify the rule, permitted action and responsible person. If two readers reach incompatible conclusions, refine the wording before asking for approval.
Keep a change log. An owner should know when to review the guideline: a new tool capability, changed data use, a reported incident or a scheduled institutional review. Avoid an unmaintained list of product names presented as permanently safe. Approval should relate to a defined service configuration and purpose, not just a brand.
Practice: prepare a one-page proposal in 20 minutes
Spend five minutes choosing one workflow and its data boundary. Spend five minutes writing allowed actions, required checks and prohibited actions. Use five minutes to test three scenarios. Spend five minutes assigning owners and marking unresolved questions for approval.
Example decision table in prose: public worksheet draft—allowed after teacher verification; named student feedback—requires the appropriate approved service and data permissions; automatic final-grade publication—not authorised by this preparation guideline. This table is useful because each result follows from a stated boundary, not from a vague judgement that a task feels responsible.
Self-review: can a reader make a practical decision from the wording? Is every consequential action attached to an accountable role? Are unresolved approvals visible, with no claim that the draft is already institutional policy? A complete proposal can contain clearly identified decisions for authorised reviewers; it must not silently invent those decisions.
Reusable prompt
Select and copy this template, or customise your own using the free builder.
Draft a proposal for AI guidance covering this workflow: [workflow without sensitive details]. Separate scope, permitted data, approved-service decisions, human checks, assessment disclosure, access alternatives, incident route and review owner. Test three hypothetical scenarios. Mark local legal, security or academic decisions for authorised review; do not invent approvals or present the proposal as adopted policy.Open the prompt builder →
Pause · Think · Check
Test your understanding
Choose one answer per question, then check your reasoning. This is practice; scores are not saved.