04 / AI-resilient assessment
Assessment that still means something when AI is in the room.
AI-resilient assessment preserves the validity of evidence about student learning even when students have access to generative AI. It aligns AI permissions with intended learning outcomes and combines final work, process artifacts, contextual application, explanation or defense, transparent AI-use disclosure, rubric judgment, and faculty decision.
01The formulation
Product + process + explanation
Not process over product. The finished work still matters; it simply stops carrying the outcome claim alone.
Product
The submitted artifact remains the centre of the assessment — it is one link in the chain, not the whole of it.
Process
Planning, drafting, annotation, and revision decisions show how the work came about, and are marked rather than merely collected.
Explanation
A short account or defense of specific choices connects the student to the reasoning in their own work.
What resilience is not
It is not AI-proofing, not detection, and not an attempt to exploit temporary weaknesses in current models.
02AI-use modes
Say what is permitted, precisely
Founding pilot scope
Most integrity disputes begin with an unstated expectation. Restricted, limited, disclosed, and required AI use can all be valid — what matters is that the mode is declared and aligned to the intended outcome.
- RestrictedNo generative AI use
- The task is completed without generative tools. Defensible only where the conditions of completion are genuinely controlled and the outcome requires it.
- LimitedPermitted for named steps
- AI is allowed for specific stages — idea generation, outlining, or feedback on a draft — while the reasoning that carries the outcome remains the student's own.
- DisclosedPermitted with a stated account
- AI-assisted work is allowed provided the student discloses what was used and where, and assessment weight shifts to revision decisions and the reasoning behind them.
- RequiredAI use is part of the outcome
- Students are expected to use AI and then evaluate, correct, or critique its output against sources, making that evaluation the assessed skill.
03Assessment Resilience Audit
A structured look at how a task fails
Founding pilot scope
The audit asks a small number of uncomfortable questions about an assignment you already run. It is service-assisted founding-pilot work, not a shipped module.
- 01
Prompt exposure
Could a competent general-purpose model produce a passing response from the task instructions alone? If yes, the task measures fluency more than learning.
- 02
Local specificity
Does the task depend on this cohort's seminar discussion, field data, placement, or set readings — material a model does not have?
- 03
Process evidence
What trace of thinking does the student leave, and is any of it assessed? Process that is collected but never marked is not evidence.
- 04
Rubric language
Do the top bands reward polished prose, or reasoning, source handling, and defensible judgment? Fluency-weighted rubrics are the easiest to satisfy artificially.
- 05
Outcome alignment
If the task were replaced entirely, would the learning outcome still be evidenced? Where the answer is no, the outcome is carried by a single fragile artefact.
Bring one vulnerable assessment
The Founding Assessment Assurance Pilot begins with an assignment you already run. Auditing where it is exposed is part of the intended scope — enrolment is not open.