Step 4 · about 30 minutes · studio
Talk with the student
You can’t see from a written product how it was made, but a short conversation can show a lot if you ask the right questions. In this step you try that. Then you look at twins: a second, smaller assessment of the same learning outcome.
§01 Time and dialogue
How the master’s thesis has always been validated
How do we get a valid picture of what students have learned in open work, without surveillance? The answer has two parts. One is time, with several points of contact with someone who follows the work, and the other is dialogue, where the student explains their own work in real time. The master’s thesis has always been validated with both, through supervision over time and an oral defence.
A short conversation about the student’s own work takes 10 to 15 minutes, with questions like these. Why did you choose this? Explain this concept in your own words. What would change if this assumption didn’t hold? Where did AI help you, and where did you overrule it? Professionals do much the same when they talk about their own work in meetings, with clients and in peer review.
Look at the AI log together with the student instead of asking about it. In one study, several students believed they had reworked the AI text a lot, while the traces showed they had barely changed it (Walton et al., 2025).
§02 Activity
Ask five questions
Three exchange students answered the same exam question as in Start. All three texts are good. The marker gave each a provisional B, and you have ten minutes with each student, enough to choose about five questions from the list. Then decide whether the student has reached the learning outcome.
The students and their answers are invented for this activity and written with AI. Each answers in good faith, as well as they can.
This activity needs JavaScript.
§03 Assessment twins
Keep the essay, give it a twin
Jasper Roe, Mike Perkins and Louie Giray (2026) would keep the essay or the take-home task and give it a twin, a second component linked to it in three ways.
Same learning outcome
Not two unrelated tasks.
Different evidence
Written and oral, or with and without AI.
Close in time
So one can be checked against the other.
Together the two give stronger evidence. You keep the vulnerable task, and if the two parts don’t match, you will see it. Writing the essay is still worth doing for what students learn, even if the essay is no longer secure evidence, and the open task goes on supporting learning while the twin confirms what was achieved.
Bolting an oral onto an essay doesn’t make a twin, the authors stress. A twin starts from a known weakness, and the whole design is built for validity from the outset. The twin itself can be a group interview, a peer conversation, a short test in class or a physical product.
Oral formats can be gamed too
A student reading out an AI-written script proves nothing. Build in live interaction, with unprepared questions and problems to solve on the spot, so that it becomes a real conversation.
Six aspects of validity
Pick the twin by asking which aspect is most at risk in your assessment. The authors use six sources of validity evidence (from Shaw and Crisp, 2012).
| Aspect | The threat | A possible twin |
|---|---|---|
| Content | AI-made work that skips the learning | A face-to-face conversation on key concepts |
| Thinking | AI takes over the intended thinking, such as the search in a literature review | The student shows or documents their own thinking |
| Scoring | High scores on AI-made content make grading unreliable | Linked scores, such as a cap on the written part if the oral is weak |
| Consequences | Surface learning and dependence on AI | Oral self-assessment and reflection on AI use |
| Generalisation | The student performs only with AI | The same outcome assessed with and without the technology |
| Link to practice | Scores say little about practice | A practice or scenario task where a person confirms competence |
How the two parts count
The parts have to be linked in grading, but not necessarily 50/50. The twin can confirm, set a threshold or count as a weighted part. In the authors’ worked example, a take-home essay is scored 0–100 and followed by a 15-minute conversation about its argument and sources.
Threshold
The student must, for example, explain at least two main arguments and answer at least one follow-up question coherently. Otherwise the essay earns nothing.
Confirming weight
The conversation is scored 0, 1 or 2, and the final score is the essay score times the conversation score divided by two. An essay at 80 with a conversation at 1 gives 40, and a conversation at 0 cancels the essay.
These models come from one assessment culture. Check how they fit your institution’s rules on grading, appeals and second examiners before you use one.
When twins fit, and when they don’t
Use a twin when
- the task is rich but vulnerable: an essay, a take-home program, a design task
- the same outcome can be shown in a lower-risk way
- a full sit-down exam is ruled out, but other checks are possible
Don’t when
- the two parts would assess different outcomes, and one redesigned task would do
- workload or fairness can’t be solved with something that scales
- a sit-down exam is feasible and already secure
By cohort size
Small (5–25)
Individual orals, group conversations with rotating student leaders, peer assessment, a conversation built into teaching.
Medium (25–75)
Group presentations, parallel group conversations with the teacher moving between them, poster sessions. Extra examiners help.
Large (75+)
A thorough twin for a random sample, or group conversations with several examiners. A sit-down exam can serve if it assesses the same outcome.
If only some students are selected, the draw has to be random and announced in advance. Suspicion and past results can play no part in it. Some systems require every student to have the conversation if it is part of the assessment, and then it has to scale. Group conversations with individual questions can do that, and so can written follow-up questions answered under supervision.
Untested, and the authors say so
The model is conceptual, and the questions it leaves open include resources, timetabling, consistency between examiners, what to do when the parts diverge, and fairness for students with other communication styles, social anxiety or another first language. Presenting twins as a hunt for misuse would cast students as cheaters and undermine trust.
Go deeper · about 15 minutes
Assessment twins. The model in more detail, with what to watch for when you pair an open task with a twin.
§04 Activity
Find the twin
For each open task, choose the component that would work as its twin.
This activity needs JavaScript.
§05 From practice
Three examples
Ex.phil. at UiT, the Arctic University of Norway
About 1,700 students a year. The graded term paper was replaced by ungraded coursework, and the grade is set in an oral exam that starts from the student’s own written work. The share who failed on campus fell from about 11 % to 4 %.
Alnes and Duarte (2026)
AI journal and conversation, Australian National University
Students build a risk plan with AI and keep a journal of their prompts, their critical judgement of the AI’s answers and their revisions. A conversation in the planning phase counts 30 %, the journal 60 % and the final report only 10 %.
Leadership, Nord University
Groups plan and teach a session for their fellow students (25 %), write a company analysis (25 %) and sit an individual exam (50 %). The peer teaching happens live in the timetable. The design was made for deep learning and turns out to be robust against AI as well.
§06 The price
What it costs
These moves come with trade-offs too, starting with time. A conversation takes around 15–30 minutes per student with documentation. Some students find oral formats demanding, and language can play a part. In a large study of oral exams, students reported anxiety, but it didn’t lead to poorer results, and afterwards more of them preferred the oral form (Davey et al., 2025). Practice conversations help, as do familiar question types and time to think.
Group projects need supervision over time, and one member can end up doing most of the work. A shared product says little about the collaboration, so a group conversation should include individual questions.
§07 Worksheet · your assessment
Sketch a conversation for your assessment
Your assessment: . The conversation lasts ten to fifteen minutes and is about the student’s own work. One good question in each row is enough, and you can leave a row empty if it doesn’t fit your subject.
What does it cost?
Sources
- Alnes and Duarte (2026) [reference to be added].
- Davey et al. (2025) [reference to be added].
- Roe, J., Perkins, M. and Giray, L. (2026). Assessment twins: an approach for strengthening assessment validity in the age of generative AI. Journal of Applied Learning & Teaching, 9(2). Open access (CC BY). doi.org/10.37074/jalt.2026.9.2.3
- Walton, J., Bearman, M., Crawford, N., Tai, J. and Boud, D. (2025). How university students work on assessment tasks with generative artificial intelligence: matters of judgement. Assessment & Evaluation in Higher Education. doi.org/10.1080/02602938.2025.2570328