Every instructor asking this question has already tried the obvious things. So let us skip past the obvious and be honest about what each option actually buys you.

Option 1: run a detector

Cheap and fast. The problem is not that detectors are imperfect, it is where the imperfection lands: published research has repeatedly found much higher false-positive rates on writing by non-native English speakers. If your class has international students, the cost of this option is paid by them, and a case built on a percentage tends to collapse the moment it is contested.

Option 2: make the assignment harder

Intuitive and backwards. Difficulty constrains the student doing the work, not the student delegating it. Every increase in difficulty raises the cost for the honest half of your class and leaves the other half unaffected.

Option 3: supervise the writing

In-class essays and invigilated exams do work, and they narrow what you can assess to what fits in a room in two hours. For many courses that is a real loss: the extended project, the lab report, the piece of code that took a fortnight. Proctoring also covers the exam window only, which is a small fraction of the assessed work in most degrees.

Option 4: ask the student about their work

The oldest option and the one that answers the question you actually have. If a student can explain their method, defend the choices, and handle a follow-up on a specific paragraph, the authorship question mostly resolves itself. If they cannot, you have something concrete rather than a suspicion.

Its real cost is time, which is why it retreated to doctoral defences and clinical exams. Five minutes per student per assessed submission is the honest number to plan around, and reviewing a recorded check runs closer to one minute once you have a routine.

The question that does most of the work

Whatever tooling you use, the shift that matters is from who wrote this to can this student account for it. The first question is unanswerable and adversarial. The second is answerable, fair to ask of everyone, and happens to be what you wanted to grade in the first place.

Do AI detectors actually work?

They produce a probability about text, not evidence about a person, and their false positives fall disproportionately on non-native English writers. Most institutions now advise against opening a case on a detector score alone.

Is it fair to make every student do an oral check?

It is considerably fairer than checking only the students who triggered a flag. A universal check is a normal part of the assessment; a targeted one is an accusation, and students experience it that way.

How long does an oral check take to run for a class of 30?

The conversations run in parallel and take about five minutes each for the student. For the instructor, reviewing thirty recorded checks with transcripts and summaries takes roughly half an hour, plus real attention on the few that look unusual.

What about students with anxiety or speech differences?

Accommodations have to exist from day one: extra time, reduced time pressure, and an instructor-run alternative that does not appear as a lesser version in any record. Delivery, fluency and accent should carry zero weight in the grade.