Every staffroom has had this conversation. A student hands in something too polished, in a register that isn't theirs, and the discussion turns to detection tools, bans, and how to prove it. That conversation almost never ends well. Detectors are unreliable, students are inventive, and the accusation is corrosive whether or not it lands.
There is a more productive version of the question: "why was this task completable by a machine in the first place?"
What AI is actually good at
Language models are strong at producing plausible, well-structured prose on any topic that is well represented in public writing. Summarise the causes of a war. Explain photosynthesis. Compare two characters. Define a term and give examples.
These are also, not coincidentally, a large share of the tasks schools have historically set because before AI they were a reasonable proxy. A student who could write a decent summary of the causes of the war had probably read about the war. The task didn't measure the summary; it measured the reading, indirectly.
AI didn't create the problem. It removed the proxy. The tasks that fall apart are the ones that were always measuring the output rather than the learning.
That is uncomfortable, but it is also the useful insight because it tells you which of your assessments are actually at risk, and it isn't all of them.
The tasks that hold up
Some assessment survives AI without any modification at all. It is worth noticing what these have in common.
Work grounded in something the model cannot access
A response to the specific discussion your class had on Tuesday. An analysis of data the students collected themselves. A reflection on the excursion, the guest speaker, the experiment that failed. The model has no access to any of it, and generic substitutes are immediately obvious.
Work where the process is the evidence
Drafts, annotated plans, revision history, a conference where the student talks through their thinking. If the assessment includes how the work developed, a finished artefact of unknown provenance is worth much less on its own.
Work performed in the room
Not necessarily an exam. A three-minute explanation to the class, a viva on a submitted piece, a practical demonstration, a whiteboard walkthrough coupled with reflection questions that demonstrate their understanding of the content. Short, low-ceremony, and very hard to outsource.
Work that requires a position and a defence of it
Models produce balanced, hedged, both-sides answers by default. A task that requires the student to commit to a judgement and to defend it against a specific counter-argument you raise, is asking for something the default output does not supply.
A worked redesign
Take a common Year 10 History task: "Explain the main causes of Federation in Australia." A model completes this in seconds, at a standard that would pass.
Now consider some alternatives, none of which require any technology to police:
- Add a source the model doesn't have. "Using the three primary sources in your class booklet, explain which cause the sources best support and which cause they underplay."
- Require a judgement. "Historians disagree about whether economic or defence arguments mattered more. Take a position and defend it against the strongest objection."
- Assess the process. Submit the plan, the annotated sources and the draft, and mark the development alongside the final piece.
- Move part of it into the room. Keep the written task, add two minutes of questions on it.
None of these are anti-AI measures. They are better history assessment. That is generally the sign you have redesigned it well: the change makes the task more valid, and harder to cheat on as a consequence.
Put your own assessment task into a chatbot and read what comes back. If the output would pass, the task was measuring something you weren't intending to measure. That is worth knowing regardless of what any student does.
Where policy still matters
None of this removes the need for a clear position on acceptable use. Students deserve to know where the line is, and it should be stated per task rather than as a blanket rule. "You may use AI to check your grammar but not to generate your argument" is enforceable and comprehensible in a way that "no AI" is not.
What redesign does is change what the policy has to carry. If the assessment holds up, the policy is a statement of professional expectations. If the assessment is fragile, the policy is the only thing standing between you and a stack of indistinguishable submissions, and it will not hold.
The part that is genuinely hard
Redesigning assessment takes time that nobody has, and it cannot be done across a whole faculty in a term. The realistic approach is to start with the assessments that carry the most weight and are the most exposed, usually the take-home written tasks in senior subjects, and leave the rest until you get to them.
It is also worth saying plainly: some very good assessment is now harder to run than it was three years ago. That is a real loss and it is reasonable to be annoyed about it. But the tasks that AI has made untenable were, on the whole, the ones where we were inferring learning from a product rather than observing it, and the redesigns that fix them are the ones most departments already had on a list somewhere marked "when we have time".
Update, September 2026: NSW has since limited Stage 6 courses to one take-home task worth at most 15%, from Term 4 2026. We cover the rule, the exempt courses and NESA's advice on detection tools in NSW has capped take-home assessment. What should other states do?