Oral Defenses and Viva Voce: A Low-Tech Way to Verify Authorship Without Detectors

A student turns in a clean, well-argued essay. The AI detector lights up at 91%. Now what? You can’t grade a number, and you can’t accuse a student based on a probability. But you can ask them to talk to you for five minutes about the thing they supposedly wrote.

That conversation has a name when it’s formalized: a viva voce, an oral defense. And it might be the most honest authorship check you have, because it skips the guessing entirely and goes straight to the question you actually care about: does this student understand their own work?

Key takeaways

• An oral defense asks a student to talk through their own work for a few minutes. It checks understanding, which is the thing you actually care about, instead of guessing at a probability score.
• AI detectors give probabilities, not proof. A short conversation gives you something you can defend in front of a student, a parent, or a dean.
• You don’t need a 30-minute thesis viva. A 3-to-5-minute targeted check, used as a spot check, scales to large classes.
• Build it into the syllabus from day one so it reads as a normal part of assessment, not an accusation.
• Pair the oral check with process artifacts (drafts, outlines, version history) and you rarely need a detector at all.

What a viva voce actually is

“Viva voce” is Latin for “with the living voice.” In doctoral programs it’s the oral defense, where a candidate sits across from examiners and answers for every claim in their dissertation. The format scales down well. Strip out the formality and you’re left with a simple idea: the person who did the work can explain the work.

That’s the whole mechanism. Authorship leaves fingerprints in a person’s head, not just on the page. Someone who wrote a paragraph can tell you why they cut the previous version of it. Someone who only pasted it usually can’t.

This matters because detection software was never built to answer the authorship question directly. It estimates how statistically “predictable” a piece of text looks and turns that into a score. The number is a likelihood, not a confession, which is why it’s worth keeping in mind that detectors output a probability, not proof. An oral check sidesteps the probability problem because you’re no longer inferring authorship from the text. You’re watching the author think.

Why the conversation beats the score

A detector tells you a sentence is unusual. It can’t tell you why. Maybe the student used AI. Maybe English isn’t their first language and they write in a more formal, hedged register. Maybe they’re just a tidy writer. A 2023 study from Stanford researchers (Liang et al., published in *Patterns*) found that several popular detectors classified more than half of a set of non-native English TOEFL essays as AI-generated, while flagging US-student essays far less often. The tool can’t see the human; it only sees the pattern.

The conversation sees the human. When a student explains, in their own words, why they structured an argument a certain way, you learn something a percentage can never give you. You learn whether they own the ideas.

There’s also the credibility problem with detectors generally. OpenAI launched an AI-text classifier in early 2023 and quietly retired it that July, citing a low rate of accuracy. When the company building the models can’t reliably catch its own model’s output, leaning on a third-party score as if it were evidence is shaky ground. A documented, fair, in-person check is ground you can stand on.

A quick mini-scenario

Picture a first-year composition class. Maya submits an analysis of a poem, and the language is noticeably more polished than her in-class writing. Instead of running it through a detector and stewing over the result, the instructor uses the oral check that’s listed in the syllabus.

Three questions, four minutes:

1. “You argue the poem’s turn happens in the third stanza. What made you pick that line over the obvious one in stanza four?”
2. “You cite a critic here. Where’d you find that source, and what made you trust it?”
3. “If you had another week, what’s the one thing you’d push further?”

If Maya wrote it, she has answers. They might be rough, she might ramble, but they’ll connect to her actual paper. If she didn’t, the cracks show fast: she defends a reading that isn’t in her text, or she can’t say where the critic came from. Either way, the instructor now has a real basis for a conversation, not a screenshot of a score.

Notice the questions aren’t gotchas. They’re about *choices*. Choices are where authorship lives.

How to run one without it feeling like a trial

The fastest way to ruin this tool is to deploy it as an ambush. Here’s how to keep it fair and low-drama.

Announce it in the syllabus. On day one, state that any assignment may be selected for a short authorship conversation, and that these are a normal part of how the course assesses learning. No surprises, no implied accusation.

Keep routine checks short. Three to five minutes. You are not re-grading the essay out loud. You’re sampling understanding.

Let students use their own work. They should be able to look at their own draft, notes, and sources while they talk. You’re verifying authorship, not testing memorization.

Ask about decisions, not facts. “Why did you organize it this way?” beats “What’s the thesis?” Anyone can reread a thesis off the page. Only the author can explain why they buried the counterargument in paragraph five.

Grade understanding, not polish. A nervous student who stammers but clearly owns the ideas passes. A smooth talker who can’t engage with their own specific text is the one to look at twice.

A simple three-tier prompt set

Build a small bank you can pull from so you’re not improvising under pressure:

• Origin: “Walk me through how this assignment started for you. What was your first idea?”
• Choice: “You did X here instead of the obvious Y. Why?”
• Extension: “What would you change or add with more time?”

The extension question is quietly the most revealing. It’s hard to fake forward momentum on an argument you didn’t build.

Scaling it to a real class size

Five minutes times 120 students is a lot of office hours. So don’t check everyone, every time.

• Spot-check randomly. Announce that a set percentage of submissions per assignment get an oral check, selected at random. Random selection keeps it fair and removes the “you targeted me” complaint.
• Weight by stakes. Skip checks on low-stakes work; reserve them for the midterm essay, the final project, the assignments that actually move the grade.
• Batch them. Run quick checks during a lab session, a workshop day, or the last ten minutes of class while others peer-review.
• Go async when you must. A 90-second video response to one targeted question, recorded by the student, isn’t a true viva but it’s a reasonable middle ground for big online courses.

The goal isn’t to interrogate everyone. It’s to make authorship verifiable and known to be verifiable. That alone shifts behavior.

Pair it with process, and you barely need detectors

An oral check is even stronger when it’s not the only signal. If your assignment already captures the *process*, the conversation almost runs itself. Outlines, annotated drafts, version history in a shared doc, a short reflection memo: these give you concrete things to ask about. “Your second draft dropped this whole section. Tell me about that” is a question only the writer can answer well.

This is also where knowing what detectors do and don’t do helps you set the right policy. If you want a low-stakes signal for *where to look*, you can still run text through a checker, but treat the result as a flashlight, not a gavel. For more on building fair, detector-light assessment, there’s more from the PaperBleach blog, and if a tool ends up part of your workflow you can compare plans on the pricing page before committing to anything.

Frequently asked questions

Isn’t an oral defense unfair to students with anxiety or language barriers? It can be if you run it like an interrogation, so don’t. Give everyone the questions’ general shape in advance, allow notes, let students point at their own draft while they talk, and grade understanding rather than fluency. A nervous student who wrote their own paper can still tell you why they chose a source. Offer a written follow-up as an accommodation. The point is to confirm authorship, not to test public-speaking nerve or accent.

How long does an oral defense need to be? For a routine assignment, three to five minutes per student is plenty. You ask two or three pointed questions about choices the student made and listen for whether the answers connect to the actual document. Reserve longer sessions for capstones and theses.

Can I require an oral defense only for students I suspect used AI? Singling out individuals invites bias complaints and feels like an accusation before any conversation happens. It’s fairer to make oral checks a normal, announced part of the course, then apply them as random spot checks or to everyone above a certain assignment weight.

What if a student clearly wrote it but just freezes up? Slow down and shift the question. Ask them to read a sentence from their own paper and explain what they meant, or to walk you through their notes. Freezing on a broad question is normal; being unable to engage with their own specific text after several tries is the signal worth weighing.

Does this replace AI detectors entirely? For most classroom purposes, a conversation plus process artifacts does the job a detector can’t, because it verifies understanding instead of estimating it. Detectors can still be a low-stakes signal, as long as you treat the number as a probability and never as a verdict.

The bottom line

You don’t need better software to answer the authorship question. You need five minutes and the right three questions. An oral defense moves the conversation from “this score looks suspicious” to “tell me about your work,” and that’s a place both you and your student can stand. Build it into your syllabus, keep it short and fair, and lean on process artifacts to do the rest.

And when you do want a quick, honest read on how a piece of writing scores, before you ever raise it with a student, check how a draft reads and treat the result for what it is: a starting point, not a sentence.