Skip to content
PlagiatScanner.de
Higher education practice · action · HSP-01

Build an AI policy for a course: components, not slogans

In brief: Begin with learning outcomes and the assessed task, not a list of product names. Decide for each activity whether AI is permitted, limited or excluded; identify which human capability must remain observable and what proportionate record is needed. Add privacy, accessibility, source checking, policy versioning and a fair clarification route. Communicate the rule before the relevant task and illustrate it with realistic examples.

Policy toolkit with decision pointsReviewed 11 September 2026

Name the learning outcome and object of assessment first

A usable rule tells students which capability is being assessed. Is it research, disciplinary reasoning, programming, written expression, reflection or a combination? The same system may be helpful during an ideas exercise and inappropriate when independent composition is the assessed outcome. Describe expected performance as observable activities.

Check alignment with the module specification, academic regulations and institution-wide rules. A course notice cannot silently rewrite higher-level authority. Involve assessment leads, privacy specialists, the library, accessibility services and student representatives where their remit is affected. Record decisions and unresolved issues.

Authority asset: AI policy toolkit with decision points

Components of an activity-based course rule
ComponentDecision questionVisible ruleControl
OutcomeWhich capability is assessed?Specific independent contributionMatches assessment rubric
ActivityWhat may AI be used for?Examples: allowed, limited, excludedAvoid vague product lists
DocumentationWhat must be traceable?Purpose, extent, selection, checkingProportionate and accessible
SourcesHow are claims verified?Original sources, not model outputHuman responsibility clear
DataWhat may be transferred?Confidential material excludedName approved environments
FairnessWhat access and language needs exist?Equivalent alternativeNo compulsory purchase
ClarificationHow is uncertainty handled?Contact, review, opportunity to respondNo automatic sanction

Classify activities rather than brands

Products and embedded features change. A durable policy describes actions: brainstorming, developing search terms, summarising sources, generating prose, language editing, explaining code or analysing data. Assign each activity a level and connect that decision to the learning outcome. “AI allowed” does not explain whether a whole submission may be generated; a blanket ban may accidentally cover accessibility tools and routine assistance.

Add boundary examples. May a system polish a student-written paragraph? Can it explain an error message? How should embedded features in word processors and search engines be handled? Examples cannot be exhaustive, but they reveal the decision rule. State what students should do when a new function appears: ask before use, document it or choose an approved alternative.

Distinguish learning activity from summative assessment. Exploratory use in a workshop can coexist with restrictions in a task measuring unaided capability. Announce the transition explicitly. A mid-task policy change needs a controlled route so expectations are not shifted retrospectively.

Require documentation that answers the assessment question

Do not default to complete conversation transcripts. They can expose personal, confidential or copyrighted material far beyond the learning outcome. A structured declaration may be sufficient: service, displayed version, purpose, affected section, form of adoption and human checking. Selected prompts, diffs or tests may be more relevant for methods and programming work.

Explain source verification. A model response is not a substitute reference. Students must open recommended literature and cite the original under the required style. Fabricated or inaccessible references must not enter the work. The AI use log template provides an adaptable structure but does not override course requirements.

Do not promise that a detector score establishes use. Automated classifiers have context-dependent error modes. A concern requires multiple sources, locatable evidence, human review and an opportunity for the student to respond. The policy should define responsibility and process rather than an automatic percentage threshold.

Design for privacy, access and revision

Do not mandate a platform until contractual terms, data flow, age requirements, accessibility and equitable access are resolved. Unpublished research, health information, assessment material and third-party data must not enter an external service without proper approval. Provide an equivalent route where an account, payment or data transfer cannot be required.

Version the policy. Each state needs scope, publication date, responsible role and change record. Link it directly from the task and rubric. Collect questions because repeated confusion identifies missing language. Revise examples before the next delivery rather than silently changing completed assessment expectations.

Approval check before publishing the course rule

  • Learning outcome and independent contribution are concrete.
  • Allowed, limited and excluded activities have examples.
  • Documentation is necessary, privacy-aware and accessible.
  • Source checking and human responsibility are explicit.
  • No detector number triggers an automatic consequence.
  • Clarification route, responsibility and policy version are visible.
  • Students receive the rule through the designated channel in advance.

The higher education practice hub connects implementation topics. The multi-source workflow for educators develops the clarification route. The one foundation connection is academic writing and evidence.

Test the draft against real task types before release. Ask reviewers who did not write it to classify plausible cases and explain their reading. Where reasonable readers reach opposing outcomes, improve the decision language. Record the exercise as policy quality control, not as a claimed accuracy experiment.

Official foundations

  1. UNESCO: Guidance for generative AI in education and research – human-centred, privacy-aware and inclusive design.
  2. German Standing Conference: recommendations on AI in educational processes – competence- and task-oriented policy principles.
  3. German Research Foundation: Guidelines for Safeguarding Good Research Practice – responsibility and traceability.

Sources reviewed 4 September 2026. Institution- and assessment-specific regulations remain controlling.