Teaching · Practical guide
Design Assessments That Reveal Learning in the AI Era
Replace promises of “AI-proof” tasks with aligned evidence of learning, process checkpoints and proportionate follow-up questions.
Begin with the learning claim
An assessment should support a claim about what a learner can do. “Produce a polished report” and “independently interpret data” are not the same claim. Write the intended skill as an observable action, then ask what evidence would show it even when writing assistance is available.
No wording makes every take-home task immune to AI assistance. Instead of treating resistance as a guarantee, design opportunities to inspect reasoning, decisions and application. TEQSA discusses assessment reform at a programme level; a single redesigned task should also fit the course’s overall evidence of achievement rather than carrying every assurance requirement alone.
Choose evidence that matches the skill
For calculation, require an interpretable method and units. For research, require source-selection reasons and a traceable claim. For a practical procedure, observe or review the procedure through an approved format. A long reflective paragraph is not automatically better evidence than a short demonstration.
Use checkpoints where they serve learning: a proposal, a small draft, a source note or a rehearsal can expose misconceptions while revision is still possible. Explain what is assessed at each stage and avoid creating unnecessary surveillance or paperwork. Protect access and fairness by providing equivalent arrangements where required.
Worked example: redesign a generic report task
Weak task: “Write 1,000 words about packet loss.” It invites a broad explanation but does not clearly show whether the learner can calculate or interpret a measurement. Revised fictional task: “Using the supplied dataset, calculate loss for each run, explain the denominator, compare two runs and identify a conclusion the data cannot support.”
Dataset: Run A sends 250 packets and loses 10; Run B sends 400 and loses 12. The correct rates are 4% and 3%. B loses more packets by count but a smaller share of sent packets. The data do not by themselves establish that a changed setting caused the difference; measurement conditions and other variables matter.
Submit a calculation table, a short interpretation and one annotated correction to an initial draft. In a brief follow-up, ask “What would change if B lost 20 packets?” The new rate is 5%, so the rate comparison reverses. This probes application to a variation rather than memorisation of the original sentence.
State how AI assistance fits the design
If the goal is independent calculation, state that the submitted method and explanation must be produced under the specified conditions. If the goal includes evaluating AI, provide a fictional generated answer containing a deliberate error and ask students to repair it with evidence. Publish the allowed assistance and declaration requirements before students begin.
Do not punish learners for using tools that the instructions appeared to allow. Also avoid treating a short oral follow-up as an unannounced replacement exam. Tell students its purpose, likely question type and contribution to assessment, and follow the course’s approved accommodation and moderation arrangements.
Build a manageable verification plan
A teacher cannot scrutinise every sentence of every draft indefinitely. Select a small set of evidence linked directly to the outcome. For the example above, inspect the denominator, calculation, interpretation and response to the changed data. Keep a concise scoring guide so follow-up questions are applied consistently.
Test the assessment with an invented strong answer and an invented weak answer. Can the rubric distinguish a copied correct number from an explained method? Does the task accidentally reward advanced language more than quantitative reasoning? Revise those weaknesses before release. Do not describe this desk review as proof that the task prevents misconduct.
Practice: a 20-minute assessment redesign
Spend four minutes identifying one assessed skill in a current-style task. Use six minutes to select two evidence types and an assistance rule. Spend five minutes creating a small worked example with an independent answer. Use five minutes to draft a follow-up variation and check it against the same criteria.
Example answer for “summarise a policy”: request a summary, a table mapping each summary claim to the supplied policy, and a revision after one policy condition changes. If the cancellation deadline changes from 12 to 24 hours, the learner should update that condition without inventing other rules. This gives more specific evidence than asking whether the prose “sounds human”.
Self-review: does each evidence requirement test the intended skill? Are tool permissions and assessment conditions explicit? Can the workload and access arrangements be justified? Keep a note of limitations: a stronger design supports judgement, but it does not offer an absolute guarantee of independent authorship.
Reusable prompt
Select and copy this template, or customise your own using the free builder.
Redesign this fictional assessment to reveal [learning outcome]. Current task: [task]. Propose two aligned evidence types, clear AI-use boundaries, a worked example with answer, and a short follow-up variation. Keep the workload proportionate and identify accessibility or approval questions. Do not promise an AI-proof assessment or use detector scores as evidence of achievement.Open the prompt builder →
Pause · Think · Check
Test your understanding
Choose one answer per question, then check your reasoning. This is practice; scores are not saved.