Section 1
Executive Summary
Key findings from 2024–2025 research across AI-resilient assessment, pedagogy, and institutional policy in higher education.
Design over Detection
Redesign assessments rather than relying on AI detectors.
Transparency is Non-Negotiable
Students need clear, assignment-level AI use policies.
Process over Product
Assess the learning journey, not just the final output.
Authenticity is the Strongest Defense
Local, personal, real-world tasks resist AI.
Oral + Multimodal = Resilience
Pair written work with live defense or video components.
AI Literacy as a Learning Outcome
Teach critical AI use as a professional skill.
Equity Must Be Centered
Address access disparities and detection bias at every level.
Section 2
AI-Resilient Assessment Frameworks
Seven frameworks and taxonomies for redesigning assessments to work with — or resist — generative AI.
Perkins, Roe, Postma, McGaughran, Hickerson — British University Vietnam (2024)
A five-level scale replacing binary "AI allowed/banned" with granular guidance. Annotate every syllabus with AIAS levels per assignment.
No AI
Complete the task entirely without AI tools.
AI-Assisted Ideation
AI for brainstorming/outlining only; all writing is human.
AI-Assisted Editing
AI for grammar, paraphrasing of human-produced work.
AI Co-creation
AI as collaborative partner; human direction + critical evaluation required.
Full AI Integration
Focus shifts to evaluating prompt design, output evaluation, and curation.
Danny Liu et al. — University of Sydney (2024)
Four design principles for evaluating and redesigning any assessment.
| Principle | Key Question |
|---|---|
| Transparency | Are AI expectations explicit and documented? |
| Agency | Does the assessment give students meaningful choices about AI use? |
| Capability | Does it build enduring human capabilities AI can't replicate? |
| Equity | Does it avoid disadvantaging students with unequal AI access? |
Building on John Biggs — Updated by Phillip Dawson (Deakin/CRADLE) (2023–2025)
Adds a fourth alignment dimension to Biggs' classic model:
- Intended Learning Outcomes (ILOs) — rewritten for AI literacy
- Teaching/Learning Activities — include structured AI practice
- Assessment Tasks — intentional AI use/non-use aligned to ILOs
- AI Affordances & Constraints — map what AI can/can't do per ILO
| Bloom's Level | AI Capability | Assessment Implication |
|---|---|---|
| Remember | Excels | Recall-based assessments are obsolete for measuring learning |
| Understand | Performs well | Comprehension checks easily gamed; use oral/proctored formats |
| Apply | Moderate | Must contextualize to local/personal/current data |
| Analyze | Struggles with nuance | Require class-specific frameworks, peer data, local cases |
| Evaluate | Lacks genuine judgment | Require personal stance, ethical reasoning, professional context |
| Create | Generates but lacks originality | Focus on process documentation + personal integration |
Situated
Requires specific local context or real-time data.
Process-Visible
Requires documented thinking, drafts, iteration.
Relational
Requires interaction with peers, community, or stakeholders.
Building on Biggs & Collis (1982)
AI is strongest at the Multistructural level. It struggles at:
- Relational — genuine integration of ideas across contexts
- Extended Abstract — metacognitive reflection, personal theorizing, transfer to novel contexts
Phillip Dawson — CRADLE, Deakin University (2023–2025)
- The arms race is unwinnable — detection tools will never reliably work
- Threat model your assessments — systematically identify AI vulnerabilities
- Layer defenses — process evidence + oral component + authentic tasks
- Assessment conditions are design variables — supervised vs. unsupervised, timed vs. untimed
Ethan Mollick — Wharton School (2023–2025)
- "Everyone gets an AI" — design assuming universal access
- The Jagged Frontier — AI capabilities are uneven; target the gaps
- Instructors must use AI themselves before designing assessments
AI Banned
No AI use permitted at any stage.
AI as Tool
AI as a specific, bounded utility.
AI as Collaborator
Human-AI co-creation with reflection.
AI as Focus
Critical evaluation of AI IS the objective.
Section 3
Practical Activity Designs & Pedagogical Models
Concrete models and redesigned assignment examples across disciplines.
AI as Tutor
Explanations, Socratic dialogue.
AI as Tool
Brainstorming, data analysis, drafting.
AI as Subject of Study
Examining outputs, biases, limitations.
AI as Collaborator
Co-creation with critical reflection.
- Mentor Guided, ongoing feedback.
- Tutor Direct instruction with Socratic questioning.
- Coach Metacognitive reflection prompts.
- Teammate Shared task collaboration.
- Student ⭐ Student teaches the AI — triggers elaborative interrogation.
- Simulator Scenario-based learning environments.
- Tool Specific technical tasks under student direction.
The "Visible Process" Portfolio
- Timestamped iterative drafts showing progression
- Reflective metacognitive essays at each stage
- AI interaction logs annotated with reasoning
- Grading weight: 50–70% process, 30–50% product
Oral Defense / Viva Voce
- 5–10 minute structured conversations about submitted work
- "Walk me through" protocols targeting specific passages
- Video-recorded explanations for scalability
The "Unfolding Case" Model
Problem revealed in stages over time. Decisions required at each stage, justified IN CLASS. AI cannot help because information is released in real-time.
DIEP Reflective Journal
- Describe — What did I do? Did I use AI?
- Interpret — What did I learn? How did AI help/hinder?
- Evaluate — Was AI output accurate? What did I correct?
- Plan — What will I do differently?
| Principle | Low Authenticity (AI-vulnerable) | High Authenticity (AI-resilient) |
|---|---|---|
| Local | Generic/decontextualized | Locally situated, student-collected data |
| Lived | No personal connection | Requires personal experience/reflection |
| Live | Unlimited, asynchronous | Synchronous, staged, real-time |
AI-Excluded
Supervised exams, handwritten work, live performances.
AI-Resistant
Process-heavy, personalized, original data tasks.
AI-Inclusive
AI use permitted, scaffolded, documented, critically evaluated.
AI-Integrated
Learning to use AI effectively IS the objective.
| Traditional | AI-Era Redesign |
|---|---|
| 5-page argumentative essay | "AI Audit" Essay — Generate AI essay, write critical analysis of its weaknesses |
| "Dialogue with AI" — Multi-turn conversation + reflective essay on what AI missed | |
| Situated Personal Narrative + Analysis — Personal experience analyzed through theory | |
| Iterative Peer Review Workshop — Full draft cycle documented and assessed |
| Traditional | AI-Era Redesign |
|---|---|
| Problem sets, lab reports | "AI as Lab Partner" — Evaluate AI's experimental design, identify 3+ flaws |
| Error Analysis — Find and correct subtle errors in AI-generated solutions | |
| Data Storytelling with Original Data — Student-collected data + interpretation focus | |
| Code Review + Refactoring — Review AI code for efficiency, security, edge cases |
| Traditional | AI-Era Redesign |
|---|---|
| Research paper | Original Fieldwork — Interviews, ethnography in student's own community |
| Policy Brief for Local Issue — Local data + stakeholder interviews | |
| "AI Bias Audit" — Systematically examine AI outputs for cultural bias | |
| Simulation/Role-Play — Real-time decisions drawing on disciplinary knowledge |
| Traditional | AI-Era Redesign |
|---|---|
| Art history essay | Curatorial Project — Select works, write wall text, justify choices |
| "Human vs. AI" Creative Workshop — Create alongside AI, write critical comparison | |
| Site-Specific Art Response — Visit location, create response engaging with place | |
| Annotated Creative Process Portfolio — Document every stage with reflective commentary |
- AI Foundations — what AI is, how it works, capabilities/limitations
- AI Ethics & Social Impact — bias, fairness, privacy, environmental costs
- AI Application — using AI tools effectively for learning and work
- AI Evaluation — critically assessing outputs for accuracy and bias
- AI and Human Agency — knowing when to use/not use AI
17 competencies across 5 themes: What is AI? · What can AI do? · How does AI work? · How should AI be used? · How do people perceive AI?
Section 4
Institutional Policies & Quality Assurance
Positions from major quality assurance bodies, integrity frameworks, equity considerations, and competency-based education.
| Organization | Key Position |
|---|---|
| CSU System | Three-tier approach (system → campus → course); QLT supplemental guidance for AI |
| UC System | Pedagogy over policing; faculty autonomy; no blanket AI bans |
| ACM/IEEE | Assess process not product; AI as collaborator; mirror professional practice |
| AAC&U | VALUE Rubrics are inherently AI-resilient; High-Impact Practices resist AI |
| TEQSA (Australia) | Landmark "Assessment Reform for AI Age" — design spectrum: AI-proof → AI-resistant → AI-inclusive |
| QAA (UK) | Assessment design first; student partnership; diversified methods |
| ICAI | Graduated AI use policies; educate over sanction; detection tools as one data point only |
| UNESCO | Human agency; ethical use; inclusive design; critical AI literacy |
CSU QLT (Quality Learning and Teaching)
- Section 2 (Assessment): Must specify permitted/prohibited AI use
- Section 6 (Technology): AI tools under instructional technology umbrella
- Proposed AI Supplement: Explicit AI policies, AI-resilient design, critical AI evaluation
Quality Matters (QM) — 7th Edition
- Standard 3: Assessment instructions must specify AI policies
- Standard 5: Activities must state AI permission level
- Standard 6: AI tools must be accessible with instructions
- QM recommends AI Use Spectrum Statements in syllabi (Levels 0–4)
OLC Quality Scorecard
- Assessment integrity by design
- AI literacy as institutional competency
- Ethical learning analytics with AI
Old Model
Detect AI use → Punish
New Model
Design assessments where AI use is either: (a) irrelevant, (b) impractical, or (c) the actual learning objective.
| Finding | Evidence |
|---|---|
| Detection should NOT be sole evidence | ICAI, TEQSA, QAA, most researchers |
| High false positives for non-native English writers | Liang et al. 2023; Columbia/Stanford research |
| Paraphrasing defeats detection | Multiple studies |
| Mixed human-AI text is unreliable to detect | Industry consensus |
| Detection is a losing long-term strategy | Dawson, Perkins, and others |
| Strategy | AI Resilience |
|---|---|
| Oral assessments / vivas | Very High |
| Process portfolios | High |
| Personalized/contextualized tasks | High |
| In-class supervised work | Very High |
| Multimodal assessments | High |
| Scaffolded multi-stage submissions | High |
| AI-inclusive with critical analysis | Moderate–High |
| Collaborative/team assessments | Moderate–High |
Lumina DQP
Emphasis on applied/collaborative learning.
WGU Model
Performance-based assessment by expert evaluators.
C-BEN Framework
Assessments must measure human capability, not just output.
NACE Career Readiness
8 competencies assessed through experiential demonstration.
Open Badges / Micro-credentials
Tied to demonstrated skills.
| Issue | Implication |
|---|---|
| Digital divide | Premium AI tools cost money; institutions should provide equitable access |
| Detection bias | AI detectors disproportionately flag non-native English speakers |
| AI tool bias | AI outputs embed training data biases; teach critical evaluation |
| Faculty workload | Assessment redesign burden falls disproportionately on contingent faculty |
| Student privacy | FERPA implications of student work in third-party AI platforms |
Section 5
Cross-Cutting Themes & Consensus Principles
Seven principles emerging across ALL frameworks, organizations, and researchers.
Design Over Detection
Assessment redesign is more effective AND more equitable than AI detection. Detection tools fail technically and disproportionately harm non-native English speakers.
Transparency and Explicitness
Every assignment needs clear, specific AI use expectations. Use scales like AIAS to give students a shared vocabulary and remove ambiguity.
Process Over Product
The most AI-resilient assessments require evidence of the learning journey: drafts, reflections, iteration logs, and oral defense.
Authenticity is the Strongest Defense
Local, personal, real-world, and embodied tasks are inherently harder to outsource. Situated + process-visible + relational = the AI-resilience layer.
Oral + Multimodal = Resilience
Pairing written work with oral defense, video, or live demonstration significantly increases integrity without requiring AI detection.
AI Literacy as a Learning Outcome
Critical, ethical AI use is an emerging graduate competency across all disciplines. Build it into your course outcomes explicitly.
Equity Must Be Centered
Address access disparities, detection bias, and faculty workload inequity at the institutional level. Equity is not an add-on — it's a design principle.
Section 6
Quick-Reference: Frameworks at a Glance
All 15 frameworks summarized in one table for fast lookup.
| Framework | Author / Org | Type | Best For |
|---|---|---|---|
| AI Assessment Scale (AIAS) | Perkins et al. | 5-level scale | Syllabus annotation, policy clarity |
| TACE Framework | Liu / U. Sydney | Design principles | Assessment review and redesign |
| Constructive Alignment 2.0 | Biggs / Dawson | Alignment model | Curriculum-level redesign |
| Bloom's + AI Mapping | Various | Taxonomy update | Individual assignment design |
| SOLO + AI | Biggs & Collis | Taxonomy | Rubric design, outcomes writing |
| Defence in Depth | Dawson / CRADLE | Security model | Assessment security audits |
| Jagged Frontier | Mollick / Wharton | Practical pedagogy | Faculty development |
| Local / Lived / Live | Various | Design heuristic | Quick assignment redesign |
| Process Portfolio | Eaton et al. | Assessment method | Written assignments |
| AI as Interlocutor | Furze, Mollick | Assessment design | Advanced/capstone courses |
| Harvard AI Pedagogy | Harvard metaLAB | Activity database | Cross-disciplinary activities |
| 7 AI Roles | Mollick & Mollick | Pedagogical model | Classroom AI integration |
| UNESCO AI Competency | UNESCO | Literacy framework | Program-level AI literacy |
| TEQSA Assessment Reform | TEQSA (Australia) | Policy framework | Institutional assessment policy |
| VALUE Rubrics | AAC&U | Competency rubrics | Authentic, developmental assessment |
Section 7
Recommended Next Steps
Prioritized actions for immediate use, course redesign, and institutional/program level.
Adopt AIAS
Tag every assignment with an AI Assessment Scale level. Gives students clarity and builds shared vocabulary.
Apply the "Local, Lived, Live" Test
Audit existing assignments against these three qualities. Any that fail all three need redesign.
Add Oral Components
Pair major written assignments with 5–10 minute structured conversations. Dramatically increases integrity with minimal grading burden.
Stress-Test Your Prompts
Run every assessment through AI before assigning it. If it earns a B+, redesign before the semester starts.
Use TACE as a Design Checklist
Review each assessment through all four lenses: Transparency, Agency, Capability, Equity.
Implement Process Portfolios
Shift grading weight toward documented process (50–70%). Timestamped drafts + reflections + AI logs.
Create an AI Use Spectrum Statement
Add to every syllabus — specify AI permissions per assignment using AIAS levels 0–5.
Design at Least One "AI as Interlocutor" Assignment
An assignment where critical evaluation of AI IS the objective. Builds AI literacy as a graduate competency.
Extend QLT Reviews with AI-Specific Criteria
Add supplemental AI criteria to existing QLT course review processes.
Add "AI Resilience Level" to Alignment Matrices
Include as a column in program-level outcomes alignment documentation.
Invest in Faculty Development
Educators must use AI themselves before they can design for it.
Ensure Equitable AI Access
Institutional licenses so no student is disadvantaged by inability to pay for premium access.
Section 8
Key Resources & URLs
Essential links for further reading, tools, and implementation guides.