Assessment in the age of AI

You set a short essay for homework, and a good number come back fluent, tidy, and strangely hollow: correct in outline, empty of the student. You can't prove any of it was written by a chatbot, and that is the point. Assessment did not break because students started using AI. It broke because a lot of what we were setting only ever asked them to reproduce, and reproduction is now free. The useful response is not to police the tool. It is to set work that is worth doing again.

Start by giving up on detection

The instinct is to reach for an AI detector, and it is the wrong first move. Detectors are unreliable, they change their verdict on the same text from one week to the next, and their false positives land hardest on the students least able to argue back, including those who write in a second language or simply write plainly. Accusing a student on a bad signal does more damage than the essay ever could.

So assume you cannot tell from the finished text alone. That sounds like a loss, but it clears the ground. Once you stop trying to catch the tool, you can start designing tasks where using it does not get the student very far in the first place.

The real question AI changed

A chatbot is genuinely good at a narrow, important thing: producing fluent, generic explanation. Summarise a topic, define a term, walk through a standard method, write a competent five-paragraph essay on a common prompt. If a task only asks for that, it now measures almost nothing about the person handing it in.

What AI cannot do well is reason from something it was never given: your student's own data, the argument your class actually had on Tuesday, a local case, a specific text you annotated together, the working behind a wrong answer. That gap is the whole design brief. Move your tasks toward what is particular to this student and this room, and away from what is generic to the topic.

Anchor the task to something only your student has

The single most effective change is to require material a chatbot does not have. Ask students to build on their own experiment results, respond to a discussion from your class, apply a concept to an example from their own town or placement, or analyse a source you studied together rather than a famous one. The more the answer depends on being in the room, the less a general-purpose tool can produce it for them.

Assess the process, not just the product

A finished answer hides how it was made. Ask to see the making: an outline before the draft, the two sources they rejected and why, a short note on the decision they found hardest, a marked-up first attempt. Process work is far harder to fake convincingly than a polished result, and it also happens to be where the real learning is. You are not adding surveillance, you are marking the thinking instead of the output.

Bring the high-stakes judgement back into the room

Not everything needs to be a locked exam, and making everything one would gut the good that take-home work does. But for the few moments where the grade really matters, put the thinking somewhere you can see it: a short piece of in-class writing, a two-minute oral where the student explains their own submission, a viva on the coursework they handed in. A student who understands their work can talk about it. A student who outsourced it cannot, and you will know within a question or two.

Turn the AI into the thing being assessed

Instead of pretending the tool does not exist, put it on the table. Give the class an AI-written answer and ask them to find where it is wrong, thin, or confidently mistaken. Ask them to improve it and justify each change. Evaluating an answer is a higher-order skill than producing one, it is exactly what you want them to be good at, and it is a genuinely useful thing to be able to do in a world full of plausible machine text.

What not to do

• Don't lean on detectors as evidence. Treat a flag as a prompt to look closer, never as proof.
• Don't ban AI outright across the board. It is unenforceable, and it wastes the chance to teach judgement.
• Don't turn every task into a closed, timed exam. You would trade a real problem for a worse one: less feedback, more anxiety, shallower learning.
• Don't accuse on a hunch. If integrity is in question, talk to the student about their work and let their understanding speak.

The fundamentals did not change

Strip away the panic and assessment is asking what it always asked: does this student understand the material, and can they do something with it. AI did not end that. It ended the assessments that were never really answering it, the ones that rewarded neat reproduction and called it learning. The work now is to ask the question more honestly, with tasks rooted in your own classroom and judged partly on the thinking, not only the tidy result.

There is an honest workload cost to this. Writing fresh, applied questions every term, building enough variety that answers can't just circulate, running quick in-class checks, all of it takes time you do not have in surplus. That is the practical reason tools like XAM exist: to take the mechanical part of producing varied, level-appropriate questions off your desk, so the effort goes into the design choices that actually matter. The judgement stays yours. The typing does not have to.