The test
Did the AI do the thinking that the assignment was assessing? That's it. That's the line, and it works across almost every case.
An essay is assessing whether *you* can construct an argument. If the AI constructed it, the assessment measured nothing and you have submitted a forgery. A problem set is assessing whether you can apply a method. If the AI applied it, same. But if the assignment assesses your argument and the AI helped you find a typo, nothing was faked — the thing being measured is still yours.
Where it falls, in practice
| Use | Assessed thinking done by | Verdict |
|---|---|---|
| Generating the essay you submit | The AI | Misconduct |
| Generating code you submit as yours | The AI | Misconduct |
| Asking it to solve a graded problem, copying it | The AI | Misconduct |
| Asking it to explain a concept you don't get | You, afterwards | Fine |
| Asking it to find the error in *your* attempt | You | Fine |
| Asking it to attack your thesis | You | Fine, and useful |
| Using it to generate practice questions | You | Fine |
| Proofreading your own prose | You | Usually fine — check |
Two things people get wrong
"I rewrote it in my own words, so it's fine." No. Paraphrasing generated work still means the *thinking* was generated. The words were never what was being assessed.
"Detectors don't work, so I won't be caught." AI detectors are indeed unreliable — and this is a bad reason to feel safe. People are caught by supervisors noticing the work doesn't match how the student talks, by vivas, by fabricated citations that are trivially checkable, and by the gap becoming visible in a closed-book exam. Betting on a detector's false-negative rate is a strange thing to build your degree on.
And the reason that isn't about rules
The point of a problem set is not the problem set. It is what happens in your head while you do it. Outsource that and you have completed a task and skipped an education — and you find out in an exam hall, some months later, with no laptop.