MOODLE · ACADEMIC INTEGRITY
AI Oral Exams in Moodle: Verify Authorship Without an AI Detector
Detection answers unreliably. Verification asks whether the student can account for the work. How to move your integrity strategy to the second question, inside Moodle.
By Eduface · Updated July 2026 · 11 min read
Detection asks whether a text was written by a machine, and answers unreliably. Verification asks whether the student can account for the work, and answers reliably. This article is about moving your integrity strategy from the first question to the second, inside Moodle.
The short answer
AI text detectors cannot support a misconduct finding, because they produce false positives on legitimate student writing and cannot be explained to a panel. Oral assessment verifies authorship directly by asking the student to defend the submitted work in a structured conversation. AI conducted oral assessment makes this affordable at cohort scale, where a human viva for every student is not. The academic decision stays with the lecturer.
What detection measures
Statistical properties of text
What a panel needs
Evidence a student can defend
What oral assessment measures
Whether the student can account for the work
Delivery in Moodle
LTI 1.3 external tool, no plugin
Typical use
A short structured defence after submission
Who decides the outcome
The lecturer, always
Why do AI detectors fail as integrity evidence?
Three reasons, and each one alone is disqualifying.
They produce false positives on real student writing. Detectors work by measuring statistical regularity: how predictable each word is given the words before it. Text written by a language model tends to be more regular than text written by a person. So does text written by a careful student following a rigid essay structure, by a student writing in a second language, or by a student who has been taught to write plainly. The properties detectors reward are the properties good academic writing instruction produces.
The failure is not evenly distributed. The students most likely to be wrongly flagged are the ones writing in an additional language and the ones who most closely followed the guidance they were given. An integrity process that systematically disadvantages those groups will not survive scrutiny, and should not.
They cannot be explained. A misconduct panel needs a member of staff to state what the evidence shows and how it was produced. A probability score from a proprietary classifier that cannot be interrogated is not that. Any decent defence will ask what the false positive rate is on this cohort, and there is no good answer.
None of this means the underlying worry is misplaced. A written submission produced outside supervised conditions no longer evidences that the student can do the thing. That is a real problem. Detection is simply the wrong instrument for it.
What does verification look like instead?
You change the question. Instead of asking whether the text was generated, ask whether the student can account for it. This is not new. A doctoral viva has done exactly this for centuries, and no one has ever needed a detector to establish authorship of a thesis. What is new is the need to do it for four hundred undergraduates rather than one candidate, which is where it has always broken down. A ten minute oral for four hundred students is roughly 67 hours of academic time, plus scheduling, rooms and reasonable adjustments. Most departments cannot do it, so they have not. That constraint is what AI conducted oral assessment removes.
How does an AI oral exam work?
Four steps, with academic judgement at both ends and the conversation in the middle.
1. The lecturer configures the session. Three modes are available and the choice sets everything downstream.
Mode
What it is for
Integrity Check
Confirming that the student can account for work they submitted
Oral Exam
A complete oral assessment, criterion graded against your rubric
Practice Viva
Students rehearse against the real questions from an upcoming live session, as many times as they want
You then set the rubric or the question parameters and link the session to the student’s submission. You also set the examiner’s character and dialogue style, and this matters more than it sounds. Four styles are available: Socratic, which asks open ended probing questions and builds toward deeper reasoning. Confrontational, which pushes back on every claim and requires the student to defend a position. Supportive, which encourages and guides to draw the best answer out. Formal, which holds a strict academic tone and procedural structure throughout. A viva for a final year dissertation and an integrity check on a first year essay should not sound the same, and here they do not have to. The character can be set to the discipline as well as the register: a strict dissertation examiner, a sceptical investor reviewing a business plan, or a distressed patient in a simulated clinical encounter.
2. The session is published to the LMS. Eduface pushes the session link to the student through Moodle over LTI 1.3, the same connection covered in our AI Essay Grader for Moodle article. No separate login, no room booking, no scheduled slot, and nothing for IT to deploy.
3. The student takes the oral in the browser. They open the link and speak their answers. The session records automatically and typically runs 15 to 20 minutes. The questions are generated from the student’s own submitted work: their argument, their cited sources, their methodology. This is what makes it verification rather than a general knowledge test, and it is why the format is difficult to prepare for with generated material. There is no fixed question set to prepare against, because the questions come from the student’s own text and from their own previous answer.
4. The lecturer reviews, approves, and the result syncs back. The lecturer receives the full transcript and a criterion by criterion evaluation with the reasoning behind it. They form their own judgement, change anything they disagree with, and approve. Only then do the grade and the transcript sync back to the Moodle Gradebook. Where the oral raises an integrity concern, it enters your existing academic misconduct process as a transcript a member of staff can read and explain, which is precisely what a detector score is not.
Where does this fit in an assessment strategy?
Three patterns work, in increasing order of cost.
As a checkpoint on a sample. Integrity Check mode on a randomly selected proportion of the cohort, with the written submission marked as usual. The deterrent effect comes from the possibility of selection rather than from universal coverage. The lightest option, and often enough.
As a gateway on every submission. A short Integrity Check required from every student before the written mark is confirmed. More expensive in student time, and the strongest position on authorship.
As the assessment itself. Where the learning outcome is genuinely about spoken reasoning, negotiation, clinical communication or professional judgement, Oral Exam mode is not a check on the real assessment. It is the real assessment, and the written artefact becomes preparation for it.
Practice Viva mode sits alongside all three. Letting students rehearse against the real questions before a live session removes most of the fairness objection to oral assessment, which is that confident speakers are advantaged by unfamiliarity with the format rather than by knowing more. Oral verification also pairs naturally with early formative feedback on drafts; see AI feedback on Moodle assignment drafts.
What about fairness and reasonable adjustments?
Ask this before you procure, not after, and be sceptical of easy answers. Oral assessment has its own equity profile, and it is different from written assessment rather than uniformly better. Students with speech differences, anxiety disorders, or who are assessed in an additional language may be disadvantaged in ways that written work does not disadvantage them. Any deployment needs the same adjustments framework you would apply to a human viva: extra time, alternative formats, and a documented route to an equivalent assessment where the format itself is the barrier.
Two further points to settle before you deploy. The session records automatically and a full transcript is returned to the lecturer, so agree the retention period for both the recording and the transcript in your data processing agreement rather than leaving it to a default. And satisfy yourself about what the evaluation actually considers. Grading is criterion by criterion against the rubric the lecturer sets, which means it assesses the content of what a student says. Any system that scored fluency, hesitation or accent would be measuring something other than the learning outcome, and that is a question worth putting to every vendor in this space in writing.
What does the law require?
Systems used to evaluate learning outcomes are high risk under Annex III point 3 of Regulation (EU) 2024/1689. Article 14 requires effective human oversight and Article 12 requires logging. Following the Digital Omnibus amendments agreed in 2026, obligations for stand alone Annex III systems apply from 2 December 2027, with Article 50 transparency obligations applying from 2 August 2026. For oral assessment specifically, add two considerations. Students must be told they are interacting with an AI system, which sits squarely in the Article 50 transparency obligations arriving in August 2026. And emotion recognition systems carry their own restrictions, which is a further reason to be clear that an oral assessment tool evaluates the content of what a student says rather than inferring their emotional state from how they say it.
Frequently asked questions
Can Moodle detect AI written assignments?
Moodle has no native AI detection capability, and third party detectors integrated with Moodle produce false positives on legitimate student writing, disproportionately affecting students writing in an additional language. Detection scores are difficult to defend at a misconduct panel. Verifying authorship through oral assessment addresses the underlying concern more reliably.
How do you verify a student wrote their own assignment?
By asking them to account for it. A structured oral assessment based on the submitted work establishes whether the student can explain their choices, defend their reasoning and elaborate beyond what is written. Unlike detection, this produces evidence a member of staff can explain.
Can you run oral exams for a large cohort?
Not with human examiners at reasonable cost. A ten minute oral for 400 students is around 67 hours of academic time before scheduling. AI conducted orals remove the scheduling constraint, with the lecturer reviewing the evaluation rather than conducting every conversation.
Is an AI conducted oral exam fair?
It has a different equity profile from written assessment, not a uniformly better one. It requires the same reasonable adjustments framework as a human viva, and the evaluation should consider what the student says rather than how fluently they say it.
Does the AI decide whether a student has cheated?
No. The system produces a rubric aligned evaluation and a transcript. The lecturer forms the academic judgement, and any integrity concern goes into the institution’s existing misconduct process.
How long does an AI oral exam take?
Typically 15 to 20 minutes. The student opens the session link from Moodle, speaks their answers in the browser, and the session records automatically. There is no scheduled slot and no room booking.
How does it connect to Moodle?
As an LTI 1.3 external tool, launched from within the course. No plugin, no server access, and no dependency on your Moodle version. Approved grades and the full transcript sync back to the Moodle Gradebook.
Related reading
AI feedback on Moodle assignment drafts
Move from detecting AI to verifying learning
Oral Examination is in early access. If you are rethinking assessment for a module where authorship has become the real question, we would like to hear how you are approaching it.