Skip to content
PlagiatScanner.de
Higher education practice · action · HSP-07

Design AI-resilient assessment tasks

In brief: Start with the learning outcome and decide for each activity whether generative AI is prohibited, permitted or the object of assessment. Replace interchangeable product prompts with contextual decisions, meaningful intermediate evidence, source or data work and a short connected performance. Publish tool rules, criteria and documentation requirements before work begins. AI resilience does not mean technical invulnerability; it means that the intended competence remains assessable when tools are available.

Assessment-redesign matrixSources reviewed 4 September 2026

Begin with observable competence

An assessment is not educationally sound merely because it frustrates an AI tool. Define what students must demonstrate: disciplinary analysis, methodological judgement, source evaluation, calculation, design, reflection or communication. Then identify evidence that makes that competence visible. A long final product may blend several abilities and remain difficult to interpret without process context.

Do not promise an “AI-proof” task. No wording guarantees that a tool cannot contribute. Resilience comes from coherent evidence, clear rules and criteria focused on scholarly reasoning rather than prose surface. Product, process and a connected follow-up can complement one another without claiming total surveillance.

Programme regulations, approved module specifications and institutional assessment processes remain authoritative. A change of format, weighting or permitted aids may require formal approval. This guide offers design questions, not a universal university rule.

Authority asset: assessment-redesign matrix

Move from interchangeable output to assessable disciplinary work
Starting taskRiskRedesignEvidence
General essay on a standard topicGeneric, context-free outputAnalyse a local case using a supplied source set and test a counter-positionSource decisions, argument map, final paper
Literature summaryReproduction rather than evaluationCompare two conflicting studies and justify uncertaintySelection note, comparison table, reflection
Calculation with a final numberAnswer without understandingDiagnose a flawed solution and transfer the correction to a new caseWorking and short explanation
Programming productOpaque individual judgementExplain code, devise a test and diagnose an unseen defectRepository releases, tests, discussion
AI explicitly permittedUse remains invisibleCompare outputs, inspect limits and document retained functionsUse record, fact check, final decision

Context should not depend on privileged access

A defined dataset, case, laboratory observation or local question can make generic responses insufficient. Supply the necessary context early, accessibly and equally. Personal familiarity with the lecturer or accidental access to a placement must not become a hidden advantage.

Invite choices between plausible alternatives. Criteria, evidence, limitations and consequences should earn marks, not agreement with a concealed preferred view. A changed parameter or new source in a follow-up can test whether reasoning transfers.

Novelty alone is not quality. A breaking-news prompt may outpace training data but also restrict reliable source access and preparation. Prefer stable, authorised materials and identify the permitted knowledge date.

Use a few meaningful milestones

A topic pitch, source decision, method draft and reasoned revision can show development. Each milestone needs a disciplinary purpose and proportionate feedback or marking. Screenshots and keystroke capture create privacy burdens and do not establish authorship automatically.

Explain documentation before the assessment: what is submitted, format, retention and access. A compact AI-use record might name purpose, service, broad input type, retained function and student verification. Complete chat histories, credentials and third-party personal information should not be demanded routinely.

A short oral or practical follow-up stays connected to the submitted work: explain a choice, locate an error, evaluate a source or adapt a case. Standardise questions and criteria enough to support comparability. Fluency, accent or anxiety must not act as hidden originality tests.

Align rules, access and marking

A prohibition is fair only when AI and relevant boundary cases are intelligible. Spell checking, translation, search, code completion and generative drafting may be treated differently. Examples and an official clarification route reduce private, inconsistent agreements.

Where use is permitted, students need comparable access or a genuine alternative. Cost, disability access and data protection matter. Do not reward ownership of the most expensive model. Assess the stated outcome, disciplinary checks and transparent decisions.

Rubrics use observable language: quality of source review, method justification, response to contrary evidence, error correction and limitation analysis. “Sounds authentic” is not a defensible criterion. Detector scores should not become automated proof.

Test clarity before live assessment

Ask subject colleagues, students or the quality function to review the prompt. They can identify ambiguous tool rules, accidental barriers and broad discretion. A voluntary or formally authorised pilot may reveal usability problems; it does not prove a general effectiveness rate.

  1. State the learning outcome and observable evidence in one sentence.
  2. Define AI rules for each stage and boundary example.
  3. Review material access, privacy and accessibility.
  4. Calibrate the rubric against sample work without rewarding memorisation.
  5. Reduce milestones to those with a disciplinary function.
  6. After delivery, document issues and revise through the approved process.

The process portfolio structures milestones. The interim viva provides a connected performance. More resources sit in the higher education practice hub. The one foundation route is the guide to verifiable academic evidence.

Official and scholarly foundations

  1. UNESCO: Guidance for generative AI in education and research – human agency, privacy and competence.
  2. QAA: Maintaining quality and standards in the ChatGPT era – sector guidance on assessment redesign.
  3. TEQSA: Assessment reform for the age of AI – multiple evidence and institutional reform.

Sources reviewed 4 September 2026. The relevant institution's approved rules govern implementation.