When to Use Generative AI
Generative AI is most useful when it expands options or saves drafting time without removing human responsibility. Use the following questions before each task—not just when selecting a tool.
If the task affects admission, grades, progression, misconduct, discipline, student support, or other important rights, stop and follow institutional review procedures. See Responsible AI Use in Higher Education.
What is the purpose?
State the learning or operational outcome first. AI is a reasonable candidate when it helps generate alternatives, transform material you understand, or produce a draft for expert review. When students must demonstrate unaided writing, calculation, or clinical reasoning, do not let AI perform the capability being assessed. Assisted task completion is not enough to establish learning: in a high-school mathematics trial, unrestricted AI help improved practice performance but harmed subsequent unassisted results. That finding does not predict every university use.1
What happens if the output is wrong?
Match the review effort to the consequence. An awkward meeting-email draft is easy to correct. Incorrect feedback, a fabricated citation, or an unfair grade can materially affect a student.
Never treat model output as evidence by itself. For consequential tasks, the decision-maker must inspect the original work and relevant criteria and be able to justify the decision without referring to the model as the authority.
Can you verify it?
Use AI where you or another reviewer can recognise errors. Check factual claims against authoritative sources, recalculate numbers, open cited publications, and compare assessment-related suggestions with the approved learning outcomes and rubric. If reliable verification is impractical, change the workflow or do not use AI.
Are the tool and data appropriate?
Use institutionally approved tools and share only the minimum data needed. Prefer synthetic examples. Removing identifiers does not necessarily make a submission anonymous or suitable for an unapproved tool. Do not enter student records, identifiable work, grades, health or support information, private messages, or confidential research into an unapproved service.
Consent is not a substitute for institutional approval or a complete data protection assessment. Local models may reduce third-party disclosure, but they still require secure operation and an appropriate legal and pedagogical basis.
Is the activity fair and accessible?
Check cost, account requirements, accessibility, language performance, and availability in students’ locations. If AI use is required, provide access and an equivalent alternative for anyone who cannot use the tool. Do not infer a student’s capability or intent from writing style, language variety, or an AI detector score.
How will you communicate the use?
Publish permitted uses and disclosure requirements, explain why the tool is being used, and name who remains responsible. For student-facing or consequential outputs, explain the role AI played and provide a way to correct or challenge the result.
Quick examples
| Use | Examples | Minimum safeguard |
|---|---|---|
| Good starting point | Brainstorm lesson examples; reformat educator-authored material; draft low-stakes communications | Educator reviews before use |
| Context-dependent | Draft feedback; create practice questions; adapt reading level; summarise approved source material | Verify against originals, test accessibility and bias, disclose where relevant |
| Escalate or avoid | Admission, grading, misconduct detection, discipline, student-risk prediction, automated support allocation | Institutional assessment; never use model output as the sole basis |
If the task passes this checklist, start small, test it with representative examples, record failures, and define a condition that would make you stop using the workflow. The evaluation guide provides a reusable pilot record. Prohibited uses, including biometric emotion recognition in education subject to narrow exceptions, cannot be made acceptable by adding human review; see Responsible AI use.
References & Footnotes
Footnotes
-
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122 ↩