Verbal defenses effectively detect AI-generated cheating

What's this about?

People disagree about whether spoken talks can catch students who use AI to cheat. Teachers may ask students to explain work they handed in.

What supporters say

• Teachers can ask why students chose a source, idea, or step in their work.
• Follow-up questions can show when a student does not grasp their own essay.
• Clear rules, set questions, and trained staff can make oral tests more fair.
• A student who gave all work to AI may struggle with new questions.

What critics say

• Oral talks show what a student knows now, not who wrote each part.
• A student can learn enough from AI-made work to explain parts of it aloud.
• The proof does not show that oral talks alone can catch AI cheating well.
• Teachers need more than a spoken talk to know how much AI a student used.

The bottom line

Spoken talks can help teachers spot weak grasp of a task. But we do not have proof that they can, by themselves, tell when AI did the work.

Requiring students to explain their work aloud may help teachers spot gaps in understanding. But the evidence does not show that verbal defenses, on their own, can reliably identify cheating with AI.

The case for

A well-designed verbal defense can show whether a student understands the work they submitted. In a structured discussion, an examiner can ask about the assignment’s reasoning, sources and key decisions, then follow up when an answer is vague. That makes it harder for a student to rely entirely on polished prose that they cannot explain. 1

Research on oral examinations supports this more limited use. A review of structured viva exams in health-professions education found that they could be valid and reliable when schools used standard questions, trained examiners and clear scoring rubrics. Studies of undergraduate oral exams likewise suggest they can probe reasoning and reveal shallow understanding that a written assignment may conceal.

A small pilot study combining AI-detection software with viva-style questioning found that the oral discussion gave educators extra information about whether students could understand and explain their own submissions. Academic-integrity research also supports looking beyond the final text, including at students’ explanations and the process behind their work.

Verbal defenses may therefore make complete outsourcing less attractive. A student who submits AI-generated work may struggle if asked to explain why a source was chosen, defend an argument, or apply the same reasoning to a new example. The defense turns the task from simply delivering a finished essay into discussing the work under direct questioning. 2

Still, this evidence mainly shows that oral defenses can assess a student’s current understanding. It does not show that they can establish who wrote every part of an assignment, or how much AI assistance was used during drafting.

The case against

The central problem is a lack of direct evidence. Researchers have not established how accurately verbal defenses detect AI-generated cheating: there are no solid estimates of their sensitivity, specificity, false-accusation rate or performance in ordinary classrooms. The available pilot found that questioning added useful information, but it did not measure a generalizable rate of AI-cheating detection. 3

That distinction matters, especially when a school is considering an academic-misconduct finding. Evidence that an oral exam can measure learning in a health-professions setting is not the same as evidence that it can classify AI use in written coursework across subjects, age groups and institutions.

A poor oral performance may also have little to do with authorship. Anxiety, lack of preparation, the format of the exam, examiner behavior, conversational dynamics and fear about speaking English can all affect how well a student answers. Inclusive-design research also raises concerns about unequal participation and whether oral assessments work fairly for all students. A weak defense could therefore create false suspicion, while a confident one could provide false reassurance. 4

There are practical problems as well. Mandatory one-to-one defenses require scheduling, staff time, examiner training and consistent scoring. Oral-exam research suggests reliability is possible when tests are tightly structured, but that structure comes with substantial staffing demands. Findings from professional certification exams may also not transfer neatly to routine university or school coursework. 5

The broader evidence favors using several sources of information rather than relying on one test. A defense may be more meaningful alongside drafts, revision histories, source notes, in-class work and clear rules on permitted AI use. Automated AI detectors have their own well-known false positives and false negatives, including one institutional report of disproportionate classification of non-native English writers, so they too should not be treated as conclusive.

The bottom line

Verbal defenses can be useful evidence of understanding, but they are not proven AI-cheating detectors. The strongest case is for a structured, assignment-specific conversation with trained examiners, clear rubrics and appropriate support or accommodations.

Even then, the discussion can show only what a student understands at that moment. It cannot, by itself, determine how much AI was used in creating the assignment. The evidence therefore supports verbal defenses as one part of a careful integrity inquiry, not as a stand-alone basis for accusing a student of AI-generated cheating.