Six design principles
These principles drive every redesign in this kit. Each one states the principle, the reasoning behind it, and what it means in practice. Evidence trails lead to the literature notes.
1. Decide what you are assessing before you decide about AI
The AI question is a validity question in disguise: does this task still show the learning it claims to show?
Why. Constructive alignment says outcomes, activities and assessment must point at the same capability. If an AI tool can produce a passing artefact without that capability, the task’s evidence is invalid — no detection tool can fix a validity problem.
In practice. For each task, name the capability the grade certifies. If the answer is “the final document”, the document may not be the right evidence any more. If it is “the reasoning, choices and growth behind the document”, you have a redesign path. Every case redesign starts from this step (see Case A redesign).
2. Be explicit: name an AI-use level for every part of every task
Vague rules produce accidental misconduct. Students should never have to guess.
Why. Studies of student and faculty perceptions consistently find a gap: students frame AI as normal study support; faculty frame it as an authorship question (Carvajal et al. 2026). The AIAS 2.1 gives a shared vocabulary — five named levels from No AI to AI Exploration.
In practice. Set a level per stage, not per course: brainstorming may be AI Collaboration while the final defence is No AI. Show the level with an icon and a text label, and give an example disclosure statement (every student brief in this kit does).
3. Assess the process where the process is the learning
Drafts, logs, annotations and defences produce evidence AI cannot fabricate for you — because the evidence is the journey, not just the destination.
Why. The vulnerable part of most take-home work is the single final artefact. Process evidence (annotated sources, version history, an AI-use log, a short defence) makes the thinking visible.
In practice. Where the process is part of the outcome, grade it — with rubric criteria for source verification, method decisions and reflection (all three case rubrics include process criteria). Where the process is scaffolding rather than outcome, keep it low-stakes and ungraded.
4. Secure a few checks; leave practice open
You cannot secure everything without wrecking workload and student experience — so secure the few points that carry the assurance claim.
Why. The two-lane approach (University of Sydney) separates secured assessments of learning from open assessments for/as learning, and weights the open lane higher for learning. Assurance needs only a small number of valid secure checks, not a wall around every task.
In practice. Keep weekly practice AI-friendly (AI as tutor, Full AI with a log), and put the assurance weight on short secure checks — in-class quizzes, a five-minute defence, a structured Q&A (Case B is the clearest example).
5. Design for equity, access and workload — or the redesign will not last
A redesign that costs more grading time than the original, or that penalises multilingual students and students with accommodations, will be abandoned before it works.
Why. Detection-based “solutions” disproportionately harm non-native English writers (Liang et al. 2023; Weber-Wulff et al. 2023). Oral-only designs disadvantage students with communication disabilities or high anxiety unless an alternative is built in from the start. And adoption dies under grading workload at 200+ students.
In practice. Every case’s faculty notes include: grading time vs. the original, a scaling plan for 200+ students, an accommodation plan for every oral component, and a multilingual student note. AI-use levels must be satisfiable with free-tier tools wherever the level affects the mark.
6. Do not rely on detection — rely on design
Detection is the assessment-design equivalent of a smoke alarm with no fire code.
Why. Tested across 14 tools, detectors scored below 80% accuracy, degraded sharply on edited and machine-translated text, and the study concluded they should not be the basis of misconduct accusations (Weber-Wulff et al. 2023). Detectors misclassify non-native English writing as AI-generated (Liang et al. 2023). OpenAI withdrew its own classifier in July 2023 “due to its low rate of accuracy”.
In practice. This kit contains no detection workflow and recommends none. Where assurance is needed, use a secure lane (principle 4). Where AI use is uncertain, make expectations explicit (principle 2) and assess process (principle 3).
How the principles relate
| Principle | Answers the question |
|---|---|
| 1. Validity first | What is this grade a claim about? |
| 2. Explicit AI-use levels | What is allowed, where? |
| 3. Process evidence | What shows the thinking? |
| 4. Two lanes | What is secured, what is open? |
| 5. Equity + workload | Will this survive contact with 200 students? |
| 6. No detection | What do we do instead of policing? |
The decision tool turns these principles into a short guided recommendation for your own assessment.