By ‘Why did you choose that method?’
A PhD researcher has just explained a study investigating how muscle soreness affects the way people stand up from a chair.
The examiner follows up:
‘Why did you induce muscle soreness rather than recruit participants who were already experiencing pain?’
The researcher gives a brief answer. The examiner, briefed on the researcher’s work, acknowledges what it covers, identifies what is missing, and asks a harder question about how well experimentally induced soreness represents clinical pain.
This examiner is a custom GPT created by Dr Claire Melanie Boucher.
The challenge
A PhD culminates in a viva: an oral examination in which the researcher may spend three or four hours defending their thesis to internal and external examiners.
The questions cannot be predicted. Examiners might ask why the researcher selected one method over another, how they interpret a result or what they would change. A strong response requires more than remembering facts; the researcher must explain and defend decisions made across several years of work.
PhD Supervisors help students prepare, but they cannot run a mock viva at every meeting. Repeated formal practice can also add to the anxiety surrounding the examination.
Claire wanted to give doctoral researchers a way to practice for 20 minutes at a time, focusing on one part of their work without waiting until the thesis was complete.
The solution
As part of her creative approach to working with ChatGPT Edu, Claire created ‘Viva Mentor’, a custom GPT built around an individual researcher’s study.
The first pilot focuses on a project exploring the sit-to-stand movement used in exercise rehabilitation. The study examined whether visual biofeedback helped participants distribute their weight more evenly between their legs, both normally and after exercise had caused delayed-onset muscle soreness.
Claire supplied the GPT with:
- the student’s written study
- the research literature cited in it
- a set of educational principles
- detailed instructions governing its questioning and feedback and
- a template for the end-of-session report.
The GPT is instructed to work only from these materials rather than browse the web or introduce outside evidence.
A student can request a short general practice or focus on a particular area, such as their methods or discussion. The GPT begins with questions such as:
‘Can you briefly explain what your study was trying to find out and why the sit-to-stand task was useful?’
It then responds
to the answer. In Claire’s demonstration, it recognized that the response identified the study’s main comparison but asked for a fuller explanation of why the movement mattered in rehabilitation.
The next question increased the challenge:
‘Why did you choose experimentally induced muscle soreness rather than recruiting people with existing musculoskeletal pain?’
When Claire answered that she was unsure how to address a limitation, the GPT did not write the answer for her. It suggested a structure: acknowledge the limitation, explain the purpose of the experimental model, and consider how the findings might transfer to clinical populations.
The student must still identify the relevant evidence and construct the defence.
Feedback without a verdict
The Viva Mentor is prohibited from predicting whether a student would pass or receive corrections. Claire did not want an AI system making a judgement it was not qualified to make.
Instead, when the student ends the session, it produces a report recording:
- the questions asked
- strengths demonstrated in the answers
- areas requiring fuller justification
- the questions that proved most challenging
- observations about how clearly the student communicated
- suggested areas for further practice and
- prompts for the student’s own reflection.
The student can share this report with their supervisors, but Claire cannot see their conversation automatically. This keeps the practice space separate while allowing the student to choose what they bring into supervision.
How the Viva Mentor GPT will develop
Claire spent around a day and a half developing the initial GPT with help from ChatGPT, testing and refining its instructions as she worked.
In her demonstration, the tool moved from a straightforward summary question to a more difficult methodological challenge. Its final report correctly identified the main weakness in the practice answers: they described decisions but did not yet justify them in the depth expected during a viva.
The next stage is student testing. Claire plans to observe whether the GPT remains grounded in the supplied material, increases the difficulty appropriately and produces feedback that supports—not replaces—academic supervision.
She is also checking the appropriate use of uploaded research articles and unpublished student work before any wider rollout.
If the pilot succeeds, Claire plans to create separate mentors for different thesis chapters. A student could practice defending completed work during the first and second years of the PhD, then use a combined version for a full mock viva closer to submission.
Why we loved Claire’s approach to using AI at Kent
The Viva Mentor shows how academic staff can use AI creatively to give students and PhD researchers more support without placing additional demands on supervisors’ time.
Practising with an AI tool may be particularly valuable for international students whose first language is not English and who would benefit from rehearsing a viva or oral examination several times. It also gives students a private space in which to make mistakes, test their responses and build confidence away from assessors.
The custom GPT function in ChatGPT Edu is well suited to Claire’s approach. She can give it specific instructions, limit its access to the open web and ground its questions in the researcher’s own work and selected academic literature. This keeps the practice focused and reduces (but does not fully eliminate) the risk of unsupported or invented responses.
The report produced at the end of each session completes the process. It helps students recognise what they handled well, identify areas that need more practice and prepare more purposefully for their next rehearsal.
Curious about building a CustomGPT too?
In the AI team, we can support Kent staff with training materials, advice, and even a step-by-step guide to help you stress test what you build. Get in touch with us by emailing AITeam@Kent.ac.uk
This post was drafted based on a conversation transcript with with help from generative AI, then edited, reviewed and developed by a member of the AI@Kent team.