How to set an assignment a model cannot do for the student
An assignment resists a model when it asks for something the model does not have: data from the student's own measurement, a specific class, an interview, a failed lab run, or a comparison of two sources handed out in the seminar. A generic topic takes a model a minute, and detection then deals only with the aftermath. You can tell a resistant assignment by the fact that the answer cannot be written without having attended the course.
Context the model does not have
An assignment like write about the influence of social media on language is ideal for a model. An assignment that asks for an analysis of three posts the student collected and attached is not — the model knows nothing about them and will invent them.
Anything local and recent works: data from the student's own survey, field notes, an analysis of a text discussed last week, a comparison with what was said in the seminar. The more concrete the anchor, the less the model has to draw on.
Staged deliverables instead of one hand-in
When only the finished work is handed in, there is nothing to show how it came about. When an outline comes first, then a draft, then the text, a trail appears that is hard to manufacture afterwards.
A side effect is that a student who writes honestly gets feedback in time. Checking stops being police work and becomes part of teaching.
A short defence at the end
Five minutes over the submitted text says more than any score. Ask about the content: why this method, where this figure came from, what the author would change. Whoever wrote it answers without preparation.
When there is no time for a defence, written check questions about the submitted text stand in for it. DetekceGPT builds them from the text itself, so they ask about what is actually in the work.
FAQ
Is this more work for the teacher?
At the start yes, because the assignment has to be rewritten. Staged deliverables then save time on marking and on disputes at the end of term.
Does it mean banning AI?
No. A resistant assignment makes sense even where AI is allowed — what gets assessed is then what the model cannot do.
Does it work in distance learning?
Yes. A defence can be held over video and check questions about the text work just as well in writing.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detectionUploading files, longer texts and the full check history are part of the Pro plan. Upgrade to Pro