Generative AI did not create a new problem in assessment so much as expose an old one: a great deal of what universities graded measured the production of a text, not the learning behind it. When a capable language model can draft a passable essay, solve a routine problem set, or summarise a reading in seconds, the take-home task that was standard for a century no longer reliably tells you what a student knows. This is unsettling, but it is also clarifying.
This article offers a framework for redesigning assessment when students have access to AI — one that works whether you teach in a well-resourced research university or a large-enrolment public institution with limited invigilation capacity. The goal is not to win an arms race against detection tools, which you will lose, but to design assessment whose validity does not depend on students lacking a tool they will use for the rest of their careers. See our full Teaching & Assessment cluster for related work.
Why the Take-Home Essay Broke
The unsupervised written task rested on an assumption that is no longer safe: that the effort of producing the artefact was itself evidence of learning. When production is cheap and delegable, the artefact and the learning come apart. Detection software cannot reliably close the gap — false positives fall hardest on multilingual students and those with atypical writing styles, and confident accusations based on unreliable tools cause real harm. Building your integrity strategy on detection is building on sand.
The productive response is to stop treating AI as contraband and start treating it as part of the environment students work in. That reframing changes the design question from "how do I stop them using it" to "what can I ask that shows me their thinking even when they have it." That question has good answers, and most of them make assessment better on its own terms — more authentic, more transparent, more aligned to what graduates actually do.
Assessing Validity, Not Just Detection
Validity — the degree to which an assessment measures what you intend — is the concept to put at the centre. An assessment can be perfectly "secure" and still invalid, and it can permit AI use and remain valid. Start by writing down what each task is supposed to evidence: recall, application, synthesis, professional judgement, original argument. Then ask whether the task, in an AI-rich world, still produces that evidence. Many will not, and those are the ones to redesign first.
Validity thinking also protects students. If your integrity approach relies on catching cheating rather than designing tasks where cheating gains little, you will inevitably accuse some honest students and miss some dishonest ones. Frameworks such as the UK's QAA Quality Code and equivalent standards elsewhere increasingly frame integrity as a design responsibility, not only a policing one. For how this connects to programme-level standards, see our note on constructive alignment.
Designing Authentic, AI-Resistant Tasks
The most robust redesign is toward authenticity: tasks that mirror real professional work, require the student's own context, and reward judgement over text-generation. Ask students to apply a concept to a specific case they gathered, to critique an AI-generated draft and improve it, to defend choices in a short oral, or to produce work grounded in a live dataset, a placement, or their own community. AI can assist with such tasks — as it will at work — but it cannot complete them without the student's genuine engagement.
Process-visible assessment also helps: drafts, annotated bibliographies, reflective logs, and staged submissions make the trajectory of the work assessable, not just the final artefact. This shifts the incentive from outsourcing the product to doing the thinking. Design the task so that the parts a model does well are the low-value parts, and the parts a human must do — situating, deciding, defending — carry the marks.
Programmatic Assessment and Secured Points
No single task type solves this alone. A resilient design mixes methods across a module and a programme so that any weakness in one is covered by another. Pair a flexible, AI-permitted assignment that develops skills with at least one secured assessment point — a supervised in-person task, a viva, a lab practical, or an oral defence — that confirms the individual can perform without assistance. Think in terms of a portfolio of evidence gathered over time rather than a few high-stakes artefacts.
This programmatic view eases the pressure on every task to be tamper-proof. Some assessments can be open, collaborative and AI-rich because their job is to develop and give feedback; others are secured because their job is to certify. Being explicit with students about which is which — and why — reduces anxiety and gaming. AcademicStaff's assessment-mapping tool lets you see, across a whole programme, where your secured and open points fall, so you can spot a semester that rests entirely on unsecured take-home work before it becomes a problem.
Teaching With AI, Not Only Around It
Graduates will enter workplaces where AI fluency is assumed. Pretending the tools do not exist disadvantages your students and dates your curriculum. Bring AI into the learning explicitly: have students interrogate a model's output for errors, compare it against authoritative sources, prompt it well and badly, and reflect on where it helped and where it misled. This builds the critical AI literacy employers now expect while keeping human judgement in the foreground.
Modelling responsible use also sets a professional standard around disclosure and verification. Teach students to cite AI assistance where your policy requires it, to check every claim a model makes, and to own the final work. These are the same habits your discipline expects around any source. Handled this way, AI becomes a topic your teaching addresses rather than a threat it defends against, and your assessment reflects the world your students are actually entering.
Policy, Integrity and Being Fair to Students
Clear, consistent policy is the backbone of all of this. Students deserve to know, per assessment, exactly what use of AI is permitted, what must be disclosed, and how disclosure affects marking. Vague or contradictory rules across a programme produce genuine confusion that then gets punished as misconduct — an unfairness that falls hardest on students least confident in navigating institutional norms, including many international and first-generation students.
Ground your local rules in institutional and sector guidance, and apply them proportionately. An integrity system that treats a disclosed, minor use of a permitted tool the same as contract cheating loses the trust it depends on. For the governance side of setting these standards fairly across a department, see our Leadership, Service & Committees resources. The aim throughout is assessment that is valid, humane, and honest about the tools your students will spend their lives using.
The AI-Era Assessment Redesign Toolkit
A worksheet that takes any existing assignment through five redesign moves — validity check, authenticity, process-visibility, securing, and AI-use policy — plus a programme-level secured-point map template.
Want the full article?
Enter your email for free access to the rest of this guide and our TEQSA resource library.
Frequently asked questions
Should I use AI detection software to catch cheating?
Treat detection tools with great caution. They produce false positives that disproportionately affect multilingual and neurodivergent students, and their scores are not reliable evidence of misconduct. Build validity into task design instead, and never base an integrity accusation on a detector's output alone.
Can an assessment be valid if it allows AI use?
Yes. Validity is about whether an assessment measures the intended learning, not whether tools are banned. An authentic task requiring the student's own judgement, context and defence can remain fully valid even when AI assists the routine parts, just as calculators did not invalidate mathematics assessment.
How do I stop AI undermining my whole programme?
Use a programmatic mix: pair open, AI-permitted developmental tasks with at least one secured assessment point per programme — a supervised task, viva or practical — that certifies individual capability. Map where these points fall so no qualification rests entirely on unsecured take-home work.
The AcademicStaff editorial team writes practical, evidence-based guidance for university staff — drawing on sector reporting, funder guidelines and the lived administrative reality of academic work. Every guide is reviewed for accuracy against current Australian higher-education practice.