Verbal defenses effectively detect AI-generated cheating

What's this about?

People disagree about whether spoken checks can spot work made with AI.

The key question asks if these checks can show who used AI without permission.

What supporters say

• Spoken checks can tell teachers more than AI checker scores alone, since scores can wrongly flag human writing.
• Questions about choices, reasons, and new examples can show when a student does not know the work well.

What critics say

• Few studies directly test how often spoken checks correctly find cheating, so we are not sure they work well.
• Knowing a piece of work does not prove that a student wrote it without help.

How to read this

The number of points on each side does not show who is right; stronger proof matters more than more points.

The bottom line

Spoken checks may help teachers judge a student’s knowledge and work process.

But they have not shown that they can find AI cheating with dependable accuracy.

The claim is that verbal defenses can reliably expose students who submit AI-generated work without understanding it. The evidence supports a narrower conclusion: oral questioning may improve assessment, but it has not been shown to identify unauthorized AI use with dependable accuracy.

The case for

A live discussion can reveal how well a student understands the work in ways a written submission may not. Asking students to explain their choices, defend their reasoning or apply ideas to an unfamiliar example can make gaps in understanding visible. Guidance from UNESCO, Australian higher-education authorities and a peer-reviewed review recommends more interactive, process-focused and authentic assessment in response to generative AI. Oral assessment is presented as one potentially useful part of that change. (see Figure 1) 1

Verbal questioning may also be more informative than relying on an automated AI-detector score alone. Research and reviews find that detector performance varies according to the AI model, subject area, text changes, data used to test it and characteristics of the writer. Some systems have also been found to classify some non-native English writing as AI-generated at disproportionate rates. Against that background, asking about reasoning, sources and the work process can provide information that a probability-based text score cannot. 2

The value of a defense depends heavily on how it is designed. Structured, individualized questions linked to drafts, sources and key decisions are likely to be more revealing than a short, predictable presentation. Used alongside drafts, version histories and source material, an oral defense can serve as useful corroborating evidence rather than a verdict on its own.

The case against

The central limitation is that understanding a piece of work does not prove that the student wrote it unaided. A student may be able to explain AI-generated or heavily AI-revised material, particularly if they have studied it after it was produced. AI-written text can also be paraphrased or edited, while detection results change across subjects and types of editing. A successful verbal defense therefore cannot establish by itself that no unauthorized AI assistance took place. 3

More fundamentally, there is little direct evidence showing how accurately verbal defenses detect cheating. Existing guidance, reviews and conceptual studies support oral or interactive assessment as a sensible way to measure learning, but they do not report how often such defenses correctly identify unauthorized AI use, how often they wrongly accuse students or how consistently different examiners reach the same judgment. Direct validation of the method is limited. 4

Performance can also be affected by language proficiency, disability, anxiety and unequal access to preparation. Those factors may make a student appear less confident or less fluent without showing that the submitted work was produced improperly. Transparent marking criteria and other process evidence are therefore important safeguards.

The bottom line

The evidence favours a qualified, limited use of verbal defenses, not the stronger claim that they reliably identify students who submit AI-generated work without understanding it. The support is reasonably strong for using oral questioning to reveal understanding and to supplement unreliable automated detectors. It is much weaker for treating a verbal defense as an accurate authorship test.

The key uncertainty is the unresolved difference between understanding and unaided authorship. No strong body of research yet establishes the sensitivity, false-accusation rate or consistency of real-world verbal-defense procedures across subjects, students and examiners. The most defensible approach is to use carefully designed questioning together with drafts, version histories, sources and other evidence of the work process—not to treat an oral defense as a standalone finding of misconduct.