Make learning and responsibility visible
A declaration is not a generic confession form. It should show what assistance occurred, what decisions remained with the student and how outputs were checked. The design depends on the task. Translation practice, literature review and programming may attach different significance to the same tool.
State the educational purpose before choosing fields. The declaration might support source criticism, process reflection, reproducibility or compliance with a stated tools rule. Collect only information that serves that purpose. A blanket prompt archive creates privacy risk and often says little about intellectual responsibility.
Translate “AI allowed” into specific actions
Labels such as “supportive use only” are too vague. Name activities: generating ideas, reviewing structure, editing prose, explaining or generating code, discovering sources, analysing data, translating or producing images. For each, state limits and required verification.
Specify whether the declaration is a marked component, required appendix or voluntary reflection. State consequences only where an authorised, published rule provides them. A course form cannot override assessment regulations, privacy duties or accessibility requirements.
Do not copy confidential material into the form
Prompts may contain personal data, unpublished research, case details or protected text. Ask for purpose and input category by default, rather than the full wording. Where exact inputs are educationally necessary, define controlled storage, authorised access and retention.
Do not indirectly require a student to create an account with an external service where no approved alternative exists. A student-produced record can describe use without a provider export. In group work, identify who operated the tool, who checked the output and how the group approved retention.
Service history is not conclusive evidence. Accounts may be shared, exports incomplete and logs editable. The declaration supports transparency and reflection. A particular concern must follow the authorised process rather than being decided solely by a missing chat history.
Describe human checking as an action
“I checked everything” is too broad. Useful controls include opening every cited source, comparing a calculation with a known case, running code in a recorded environment, back-checking technical terms after translation or reading each edit against the original passage. The check should match the task’s risk.
For literature suggestions, verify both existence and claimed support. For analysis, retain code, parameters, environment and outputs. For language assistance, protect meaning and uncertainty. A well-designed declaration teaches these differences during the work rather than asking students to reconstruct them at submission.
The independent-work rubric keeps reflection and product quality distinct. The guide to communicating an AI score prevents detector output replacing a declaration. The higher-education practice hub joins related resources. The single foundation route is the evidence-review guide.
Test the template before graded use
Ask two staff members and a small voluntary group to complete fictional scenarios. Look for terms interpreted differently, fields that solicit unnecessary data and unreasonable workload. This pilot tests clarity and process only; it does not establish a misconduct rate or prove educational effectiveness.
Publish the rule, template, example, submission route and contact in an accessible format. Make clear that truthful disclosure of permitted use is not in itself grounds for a lower mark. Where an activity is excluded, students need the boundary and learning rationale before assessment.
- Identify learning outcomes and the local rule.
- Name permitted activities and limits.
- Select the minimal or detailed version.
- Define data-minimising fields and secure submission.
- Provide examples of human verification.
- Test fictional cases for clarity and accessibility.
- Version the template and document its period of use.
Revise after feedback or policy changes, but do not apply new fields retrospectively to completed work. A dated scope lets students and reviewers identify exactly which declaration governed the assignment.
Handle non-use statements with equal precision
If the course asks students to declare that no generative system was used, define the category first. Spell-checking, predictive text, statistical software and generative chat may be treated differently. A statement cannot be truthful if ordinary embedded tools are left undefined.
Offer a line such as: “I did not use a generative AI system for planning, drafting, analysis or revision of this submission; the following non-generative tools were used …”. Adapt it to the real rule. Do not ask students to guarantee facts about hidden software features they could not reasonably observe.