Educator guidance for AI-integrated assessment
Implementation guidance, responsible-use principles, and adaptation notes for educators applying the SAGE assessment policy.
Why a structured process matters
University policy on generative AI has progressed beyond blanket prohibition. Conditional permission, disclosure requirements, and integrity safeguards are now common. Cross-national research examining hundreds of institutions has shown that policy development remains highly decentralised, with most universities leaving AI decisions to individual instructors rather than embedding a coherent pedagogical model.
Critical reviews have further shown that institutional discourse is dominated by concern over originality. The policy problem is framed as how to protect authentic authorship, rather than how to cultivate disciplined forms of human–AI collaboration. Most universities now specify whether AI may be used and require acknowledgement when it is. Fewer provide a structured pedagogical process showing students how to work with AI responsibly. It is this underdeveloped pedagogical layer that SAGE addresses.
Determines whether AI may be used in an assessment and under what conditions.
Guides how students work with AI during task completion through a human starting point, verification, visible refinement, reflection, and defensible authorship.
Verifies disclosure, authorship, and defended understanding, and applies institutional integrity procedures where required.
SAGE and SAGE-AP
SAGE is the broader six-stage pedagogical framework. SAGE-AP is the student-facing policy instrument derived from it, translating that process into fifteen clauses and practical companion resources.
Permission remains separate
SAGE-AP does not determine whether AI is permitted. Institutional policy or a framework such as the AI Assessment Scale (AIAS) sets the boundary; SAGE-AP explains how students work responsibly within it.
How the fifteen clauses are organised
Foundational
Scope, local precedence, student authorship, explainability, and accountability.
Operational boundaries
Permitted uses, prohibited uses, task-level conditions, and practical limits.
Substantive principles
Verification, bias, disclosure, privacy, intellectual property, and the scope of AI-supported work.
Integration and assurance
The SAGE workflow, Defend, integrity thresholds, good-faith competence gaps, and final responsibility.
Evidence base
Published SAGE studies and sector recognition include implementation with more than 1,500 students across 30+ studies in Australia, China and the United Kingdom, three TEQSA Knowledge Hub listings, and reference to SAGE by institutions including the University of Warwick.
Why Defend exists
In one structured audit reported in the SAGE evidence base, only 12% of unsupervised submissions produced a genuinely traceable evidence trail. The Defend step addresses this gap through supervised verification of reasoning ownership.
Learning outcomes
SAGE makes specific competencies assessable: disciplinary judgement, evidence-based reasoning, critical evaluation of AI, reflective practice, and the ability to defend intellectual work.
Generate
AI produces initial output after a human baseline.
Evaluate
Output checked against authoritative disciplinary anchors.
Refine
Accept, modify, reject decisions with evidence.
Audit
AI critiques the work; student judges the critique.
Reflect
Student identifies AI strengths, failures, and limits.
Defend
Supervised checkpoint verifying reasoning ownership.
Responsible AI use in assessment
These eight principles are written to stand independently of SAGE. Educators may extract, adapt, and redistribute this section under appropriate attribution. Each principle includes a side-by-side comparison of responsible versus irresponsible practice.
1 Authorship and intellectual ownership
The student remains the author. AI may assist, but the final submission must reflect the student’s own understanding, analysis, and judgement. Rewriting AI output does not restore authorship if the original reasoning was outsourced.
- Student writes an outline first, then uses AI for feedback on gaps
- Student generates code with AI, then explains and modifies every function
- Student uses AI to suggest slide structure, writes all analysis independently
- Reflective content is genuinely personal; AI only checks coverage
- Student submits AI-written essay with only cosmetic edits
- Student generates entire codebase and cannot walk through it
- Student uses AI to write the spoken argument and memorises it
- Student asks AI to write a reflection on experiences they did not describe
2 Verification, accuracy, and source quality
AI output must never be accepted at face value. Students must independently verify every important claim, reference, statistic, or technical assertion. AI tools hallucinate, have training cutoffs, and draw on low-quality sources without distinction.
- AI suggests a journal reference; student locates it via the library database and confirms it exists and says what AI claims
- Student checks AI-generated statistics against the original dataset
- Student consults teaching team when unsure about AI reliability
- Student cites AI-suggested references without checking they exist
- Student accepts AI-generated case law or clinical guidelines at face value
- Student treats AI output as current without checking training cutoff dates
3 Bias awareness and critical assessment
AI tools are not neutral. They exhibit cultural bias (Western-centric training data), confirmation bias (reinforcing dominant views), and authority bias (weighting frequently cited sources regardless of quality). Gender and racial biases are also documented.
- Student evaluates whether AI output applies to an Australian legal context, not just a US one
- Student identifies that AI summary under-represents Indigenous or minority perspectives
- Student treats AI summary as a balanced account of a contested field
- Student applies US-centric AI guidance to an Australian regulatory task without checking
4 Attribution, disclosure, and the limits of acknowledgement
Three levels of attribution apply: in-text citation for quoted/paraphrased AI content; acknowledgement statement for editing or translation; and full structured declaration where detailed process evidence is required.
5 Privacy, data protection, and confidentiality
Students must not enter personal, confidential, or restricted information into public AI tools. This includes names, addresses, student IDs, thesis work, patient data, and commercial-in-confidence material. Institutionally approved platforms with data protection should be used where available. Students should also avoid using AI as a content detection tool.
SAGE: standing obligation across all steps; reinforced in formal policy Clause 86 Copyright and intellectual property
Input risk: entering copyrighted material into AI may constitute infringement. Output risk: AI-generated content may reproduce copyrighted works. Students bear responsibility on both sides.
SAGE: addressed during Evaluate and Refine (Steps 2–3)7 Embedded AI and the boundary of deliberate use
The policy applies to deliberate, task-directed use of generative AI, not to incidental autocomplete or spell-checking. However, features like Microsoft Copilot in Word or AI code generation assistants that produce substantial content do fall within scope.
8 Multimodal AI outputs
All principles apply regardless of output modality: text, code, images, diagrams, audio, video, data analysis, and design prototypes. If AI produced or substantially shaped any artefact in the submission, the same obligations apply.
Adapting the policy for your unit
The policy works best when customised for the discipline, the unit’s AI permission settings, and the format of the Defend checkpoint.
Six steps for adapting the policy to a specific assessment
1. Set the permission boundary
State whether the assessment is fully SAGE-integrated, partially integrated, or SAGE-supported with restrictions. If the institution uses an AI permission framework (e.g. AIAS), state the relevant level and note that this policy explains how to work within it.
2. Specify required anchors
Name the authoritative sources students must use in Evaluate: prescribed readings, ISO standards, legislation, clinical guidelines, case constraints. This prevents vague fact-checking.
3. Customise the evidence log
Provide a template: AI output excerpt, decision category (accept/modify/reject), justification, source used. Modify for quantitative, design, or programming tasks.
4. Define the Defend checkpoint
Specify format (viva, timed task, code walkthrough, design interrogation), duration, weighting, and whether it is individual or group-based. Challenge prompts need not be disclosed in advance.
5. Replace generic examples
Students understand boundaries best when they see them applied to a lab report, design brief, clinical case, or lesson plan from their own field.
6. Align the rubric with the process
Reward evaluation quality, refinement reasoning, audit adjudication, reflective insight, and Defend performance — not merely surface fluency.
Calibrating implementation depth
All six steps remain present in principle. What varies is the depth of evidence and formality of Defend. A formative tutorial may need only verbal discussion. A summative capstone requires the full evidence trail. Scale documentation to the stakes to prevent performative busywork.
Progressive expectations
First-year students may need scaffolded prompts and simplified logs. Capstone students should design their own prompts, select anchors, and undertake rigorous defence. Calibrate expectations to the level and stage of the learner.
Equity, accessibility, and workload
Adoption is a learning-design exercise. Evidence and assurance requirements should be proportionate, accessible, and realistic for students and teaching teams.
Manage staff workload
Prior SAGE research found that auditing AI-supported process evidence approximately doubled marking time. Use proportionate documentation and sampled, triggered, staged, or assessment-embedded assurance rather than requiring intensive review for every task.
Design for accessible participation
Even light documentation can affect students unevenly, including students using English as an additional language, students with cognitive or learning-related disabilities, and students managing significant external commitments. Provide accessible templates, clear exemplars, and tutorial support.
Avoid paid-tool advantage
Premium AI tools may provide stronger capabilities than free services. Use institutionally approved platforms where possible, and avoid assessment designs that privilege students who can afford paid tools.
Group assessment considerations
Specify: whose baseline counts; how individual contributions are distinguished; whether the log attributes AI use to specific members; and whether Defend is individual or group-based.
Defend outcome distinction
Inadequate defence in good faith is a competence matter (feedback, resubmission). Wilful misconduct (fabrication, concealment) is handled through integrity procedures. The policy should make this distinction explicit to students.
Suggested attribution
Review and currency
This policy should be reviewed at least annually or at the start of each teaching period to ensure alignment with current institutional rules, tool capabilities, and disciplinary expectations. Update the version number and effective date with each revision.