Updated September 20, 2026

Oral Assessment in the Age of ChatGPT — have them defend the work instead of detecting it

A short-defense procedure: each student defends their submitted work for four minutes under questioning. Timed, workable with 30 students, and scored on command of the content rather than on public-speaking polish.

Level: Middle school, high school, undergraduate · any subject
Duration: 55 min per defense session (8 students) · 4 sessions to cover a class of 30
Class size: Whole class, up to 32 students
Framework: Subject-agnostic. Maps to standard oral-communication and metacognition outcomes: explaining reasoning, responding to questions, evaluating the limits of one’s own work. In language courses: CEFR B1 to C1, spoken interaction.

Every staff room is asking the same question: how do you grade a written assignment when any student can obtain a competent one in forty seconds? The most common answer — buy a detector — is also the weakest.

AI detectors do not hold up as assessment instruments. Their false-positive rate is too high to support a sanction, their own vendors advise against treating the output as proof, and study after study has found that they flag writing by multilingual and non-native students far more often, precisely because that writing is more regular and uses a narrower range of vocabulary. Grading on that basis means risking a penalty against the student who writes careful, schooled prose.

There is a way out, and it is an old one: make students defend the work. A student who had their essay written for them cannot hold it up for four minutes under questioning. What they lack is the memory of doing it — what got cut, why that section moved to second, where the figure on page 2 came from, what the cited source actually says. This is not a trap. It is the difference between having done the work and having obtained it.

The procedure below is not a cheating hunt. It changes what is being graded. It fits a class of 30 without eating the term, it needs nothing but a timer and the question list further down, and it works in any subject — not only in the humanities.

Objectives

→Verify that the student actually commands the content they submitted, whatever tools produced it.
→Make the choices explicit: structure, framing, sources kept, lines of argument abandoned.
→Elicit a reasoned answer to an unrehearsed question, grounded in the submitted work.
→Practise genuine paraphrase: saying the same thing differently, for someone who has not read the assignment.
→Have the student name the limits of their own work — without that costing them marks.
→Replace suspicion, which is unverifiable and unfair, with an observation of command that you can defend to a parent or an academic integrity office.

Materials

☐The submitted written work, with at least two questions pencilled onto each paper during marking.
☐A visible timer.
☐The 20 generic questions below, handed out or posted a week in advance.
☐The 20-point rubric, one row per student, on a single sheet for the whole session.
☐Something for the rest of the class to do while defenses run: silent written work, reading, the next lesson’s exercise.

Lesson procedure · 60 min

5 min
One week out — announce the procedure, not an ambush
Teacher — Take five minutes at the end of an earlier lesson. Tell the class that submitted work will now be defended out loud for four minutes, hand out the question list, and say plainly what is being graded: command of the content, not confidence in speaking. Say what is not being graded too — you are not hunting cheaters, and using tools, AI included, is not banned as long as it is declared. This framing defuses half the conflicts you would otherwise have.
Students — Note their slot. Understand that outsourced work will still have to be defended — most adjust how they work at this point, before the first defense has even happened.
5 min
Start of the session — the room in two zones
Teacher — Set up with the papers at a table at the back, facing the queue. Give the whole class silent written work due at the end of the hour: that is the only thing standing between you and 24 spectators. Call the first student and the next one, who sits beside you and gets ready.
Students — Start the silent task. The first two called re-read their own paper — no extra notes.
4 min
The first defense — the one everyone watches
Teacher — Time it to the quarter-minute and hold the line. 60 s: “In one minute, without your paper: what is this about, and what are you arguing?” 120 s: three questions — two from the published list, one unrehearsed, drawn from your pencil marks on this specific paper. 45 s: “What is the most serious limitation of this work?” 15 s: you tick the rubric while the next student sits down. No feedback, no mark read out loud.
Students — Answer with the paper face down. The first defense sets the tone for the whole class: they see that this is neither an interrogation nor a humiliation, and that the questions are about the work, not the person.
28 min
The next seven — keep the line moving
Teacher — Run them back to back. Keep the structure identical for everyone: that is what makes comparing the marks defensible later. Vary only the third question, the paper-specific one. If a student freezes, rephrase once, then move on — freezing is not an offence, it is a rubric box. Never voice a suspicion in the room.
Students — Go in the order posted on the board. Those waiting keep working; those who are done go back to the task. Nobody leaves the room.
8 min
Two catch-up slots
Teacher — Hold two four-minute slots at the end. They cover last session’s absences, defenses that got cut short, and the cases where your rubric is still ambiguous — a student silenced by nerves and a student who does not know their own work look remarkably alike at 8:40 on a Tuesday. A second, calmer pass almost always settles it.
Students — A repeat slot carries no stigma: present it as a right, never as a summons. Any student can ask for one if they feel their defense went badly.
5 min
Close — what the mark says and does not say
Teacher — Collect the silent work. Report back on the good answers you heard, without naming anyone, and restate one last time what the rubric measures. Make clear that the oral mark does not replace the written one: it sits alongside it, and a large gap between the two opens a conversation, not an automatic penalty.
Students — Leave knowing which questions cost them marks — and therefore what to prepare for the next submission.
5 min
Right after the session — set up the rotation
Teacher — Five minutes, paper by paper, to assign the next eight and jot each one’s paper-specific question while you mark. A sustainable rhythm: one defense session a month, one assignment in three defended, with the selection drawn openly for the rest. At eight defenses per session, a class of 30 is fully covered in four sessions — roughly two hours of class time per term, with no extra marking.
Students — See the next group posted at the end of the week. Knowing you might be called is enough: the effect reaches every assignment, not just the ones that get defended.

Assessment rubric · / 20

Criterion	Points	1	2	3	4
Justifying the choices made	5	Cannot say why the work is organised this way; recites the outline or answers “it just came out like that”.	Justifies by the brief (“we had to have three sections”) with no reason specific to the topic.	Justifies the structure and at least one substantive choice with a reason tied to the topic.	Justifies the choices and can say what a different approach would have produced, and what it would have cost.
Handling unrehearsed questions	5	Freezes for an extended stretch, repeats a sentence from the paper, or answers a different question.	Answers in general terms, without connecting the answer to the content submitted.	Answers from the submitted content, with the ordinary hesitation of unrehearsed speech.	Answers, separates what they know from what they are inferring, and goes beyond what the paper contained.
Knowledge of the sources and figures cited	4	Cannot say where the sources or figures came from; plainly has not opened them.	Identifies a source but cannot say what is in it or what it contributes.	Knows what each source contributes and where it was found.	Knows what each source contributes, how much weight it carries, and why it rather than another.
Ability to restate it a different way	3	Can only explain by reciting the exact wording of the paper.	Swaps a few words; the structure and vocabulary remain those of the text.	Explains the same idea in their own words, with a different example.	Explains at two levels: a short version for someone who has read nothing, a precise one for the class.
Awareness of the work’s limits	3	Claims the work is complete; names no limitation even when prompted.	Offers a pro-forma limitation (“I could have developed it more”) with no substance.	Names a real limitation and says what it prevents them from concluding.	Names a real limitation, says what it would take to remove it, and separates what is settled from what is contested.

Ready-to-use topics

In one minute, without your paper: what is this about, and what are you arguing?
Which part took the most revision, and why?
Which idea did you abandon along the way? What killed it?
Explain your second section to someone who has read nothing, without using the words you wrote.
Where does this figure come from? What exactly does it measure?
Name a source you read and did not end up using. Why did you drop it?
Which sentence in your paper are you least sure about?
Why this structure rather than another? What would the reverse order have produced?
What is the first objection a hostile reader would raise?
Which word in your paper would you struggle to define in front of the class?
If I deleted your third section, what would collapse in your argument?
What surprised you while you were working on this?
What is the most serious limitation of what you are claiming?
What are you relying on to assert that? Is it settled, or still contested?
Walk me through the last hour before you submitted: what were you actually doing?
Which tools did you use, at what stage, and what did you keep from them?
What question does your work not answer?
What did you understand by the end that you did not understand at the start?
If you had two more pages, what would you add — and why is it not in there already?
What would you do differently if you started again on Monday morning?

Adaptations

Students with speaking anxiety

First, check that your rubric contains no criterion for posture, voice or eye contact — the one above deliberately contains none, and a shaking voice must cost nothing. Then a few accommodations that leave the procedure entirely valid: questions known in advance, a one-to-one slot outside the classroom, an explicit right to say “let me start that again” or “I don’t know, but here is how I would find out”, and the paper face down on the table, turnable once. A terrified student who, after a rephrase, explains where the figure on page 2 came from has shown you exactly what you were checking for.

Multilingual learners and non-native speakers

This is where the old approach — the detector — did the most damage, and where you need to be most careful now. Grade the content, not the syntax: full command expressed in broken English is full marks. Add 60 seconds to the speaking time, allow a glossary card of technical vocabulary, accept a word in the first language when it unblocks a sentence, and ask questions slowly, one at a time, without double negatives.

IEPs, 504 plans and other accommodations

Three variants stand up alongside the standard format. Extended time: 5:20 instead of 4:00. Written questions supplied 24 hours ahead, with only two unrehearsed questions kept for the room. Asynchronous defense: the student records answers alone, in a single take, from questions revealed at recording time. Where speaking is ruled out entirely by the accommodation, substitute a timed written interview on the same questions with no access to the submitted paper — the principle is unchanged.

Large classes (30+)

Do not defend every assignment; defend every student, at regular intervals. Eight defenses per session, one session a month, and an openly drawn selection for the assignments in between. What produces the effect is not that every paper gets checked — it is that every student knows this one might have to be defended. Above 34 students, cut to three minutes (45 s summary, two questions, 30 s limitation): you lose resolution, not validity.

Remote and online

The format survives video conferencing, provided you ask at least two questions from outside the published list and accept one honest constraint: you do not control the second screen, and you will not try to. Rely on response latency, on how a student moves from one question to the next, and on their ability to build on an answer they gave a minute ago — none of which a back-up screen supplies. Ask the student to share their draft or version history while they talk, and the conversation becomes concrete.

Group assignments

Question each member about a section they did not write. It is the only reliable way to get the whole group to read the whole document, and it settles the free-rider problem on the way. Mark individually; keep a shared mark for the coherence of the whole and nothing else.

Follow-up

Two follow-ups that make later sessions lighter. Within ten minutes of their defense, each student writes ten lines: the question they could not answer, and the answer they would give now. It is not marked; it is stapled to the paper — and it is often the most useful page in the file. From the next submission onward, require a half-page process note with every assignment: what I looked for, what I abandoned along the way, which tools I used and at which stage. It makes the defense faster, and it makes declaring AI use ordinary rather than a confession.

The same lesson with ELBO

·NOVA does not detect AI and does not claim to. What it removes is the actual obstacle: the logistics of running thirty four-minute defenses.
·Students join with a six-character code on the board — no account, no email address, no consent form to chase.
·The four-phase timer runs itself and the rubric above comes preloaded, so you tick boxes while the student talks instead of timing with one hand and writing with the other.
·The participation report keeps, for each defense, how long the student spoke and which questions were asked. That is what makes the mark hold up when a parent disputes it: you are showing an observation, not an instinct.
·A practice voice session lets anxious students run a mock defense on their own beforehand. It is the use teachers report back to us most often.
·Defenses accumulate in the student profile across years, so progress in the ability to hold up a piece of work becomes visible instead of being reassessed from zero each September.

Run it without the logistics

NOVA handles timing, speaking turns, the rubric and the participation report. One code on the board — your students never create an account.

Try NOVA free

Frequently asked questions

What if the student just memorised the text really well?

Then they did work, and that is already a win. But a memorised text does not survive the three unrehearsed questions, because those questions are not about the content: they are about choices, rejections, hesitations and sources — traces that only the making of the work leaves behind. A student who memorised an AI-written essay can deliver the first minute and will then stall on “what did you abandon along the way?”. And in the limiting case where they anticipated everything, they have done more genuine appropriation than a copied essay ever required. The procedure is not chasing purity; it moves the effort to the right place.

Is this fair, and will it hold up if challenged?

Yes, on three conditions that are not negotiable. One: the procedure is announced in advance and appears in the assessment terms given to the class. Two: it applies to everyone, not only to students you suspect — a targeted defense is an accusation in disguise and will be read as one. Three: the mark reflects what was said in those four minutes, never a suspicion about the written work. Do not penalise the same paper twice. Anything that rises to actual academic misconduct goes through your institution’s process, and only through that.

What do I do if I suspect but cannot prove it?

Do not accuse, do not give a zero, do not produce a detector screenshot. Ask your questions and write factual observations: “could not say where the figure on page 2 came from, what source 4 contains, or why section two precedes section three”. Those are observations of command — checkable, and defensible in a meeting — not an allegation. Then offer a calm conversation: ask for the draft, the version history, the reading notes. In most cases the conversation is enough and the student redoes the work. If the gap persists, the record you have built is exactly what your institution’s process needs.

How long does this actually take?

Four minutes per student, so two hours of class time for a class of 30, spread across four sessions at eight defenses each. Defending one assignment in three, that is roughly one session a month. Preparation is light: the question list is reusable as it stands, and the only thing to add is one paper-specific question, noted while you mark, which takes ten seconds. Against that, you get time back — the defense replaces part of the written feedback that most students never read.

So AI detectors are completely useless?

As a private signal telling you to reread a paper more closely, arguably not. As evidence, yes: the false-positive rate makes them unusable where an individual’s grade is at stake, the bias against non-native writers is well documented, and scores are not reproducible from one tool to the next. No serious institution rests a sanction on that, and resting yours on it exposes you. The defense has the opposite property: it produces a direct observation, made by you, that the student can contest on the merits.

Should I just ban AI?

That is the wrong question, and this procedure makes it largely moot. You can permit the tool and grade what the student did with it: ask everyone to declare which tools they used and at which stage, make it one defense question among the others, and the use comes out of hiding. A student who can say precisely what they asked an AI for, what they kept and what they rewrote is, in a real sense, demonstrating the thing you wanted.

Can I use this freely?

Yes. Print it, photocopy it, adapt the rubric to your subject, pass it around your department. It is free with no strings: no account to create, no email to leave.

Free to reproduce for classroom use. · elbo.world