Assessment twins

§01 The starting point

A question of validity

If you can’t tell whether work is the student’s own, you can’t tell whether they have met the standards of the course either. Roe, Perkins and Giray follow Dawson and colleagues (2024) in treating this as a problem of validity. Detection tools are too unreliable to base decisions on, and guidelines that only describe acceptable AI use depend on students complying (Corbin, Dawson and Liu, 2025). Responses built on surveillance may also affect the relational side of assessment, and Carless (2009) argued that reform depends on trust between students and institutions.

Our addition. This is why we don’t treat rules about AI use as a security measure. A rule says what you hope students will do, and in a task done at home nobody can see whether they do it.

The authors find structural redesign more promising, but the literature offers few concrete methods for it. They would also keep take-home essays and other unsupervised work, since essays develop abilities like self-regulated learning and metacognitive awareness that other formats can’t develop as well (Corbin et al., 2026).

Their idea of validity comes from Messick (1989). Validity is one overall judgement of how far the evidence supports the appropriateness, meaningfulness and usefulness of assessment results, and it is always made about particular tasks and their results. Each pair you design therefore has to be judged on the evidence it produces.

Validity is a property of particular assessment results rather than of the twin approach itself.

§02 Design step 1

Find the vulnerability

The authors’ definition ties a twin to a known vulnerability, such as AI completing the task or someone impersonating the student. No twin has been tried in practice yet, and they propose an iterative process in three steps, followed by a pilot. The first step takes a threshold level of AI literacy, enough to know what current models can and can’t do. Three questions help you judge the risk.

• Could AI produce high-quality work for the task with little or no input from the student?
• Can the student do it remotely, without supervision?
• How hard would it be to tell AI-generated work from the student’s own?

If most or all of these apply, a twin may help. Being unsupervised doesn’t make a task vulnerable on its own, though. The authors’ example is an art project in oils on canvas, which a student can make at home and which still isn’t meaningfully open to AI completion. For a structured check they point to the PANDORA rubric (Bannister et al., 2025), which rates how susceptible a task is to undeclared AI use.

§03 Design step 2

Start from the learning outcome

Next you spell out what the assessment has to measure, meaning what a student needs to know and be able to do in order to pass, and map both parts onto those outcomes. Reading them closely also shows which parts of the learning AI is most likely to undermine. The article gives two examples.

Real-time formats such as oral explanations and group discussions may be out of reach with large cohorts and few resources, and then the authors suggest exploring peer assessment, team-based assessment for instance. Group work can be vulnerable too, if individual contributions are produced with AI.

A twin can be low-tech, like an oral defence, or high-tech, like a concept map that a group builds with AI tools in class. For generalisation the authors suggest assessing the same outcomes in low- and high-technology settings, restricting AI where needed.

Our addition. So the twin doesn’t have to be the part without AI. Students use AI in several roles, working alone, letting AI help or handing parts of the work over, and they move between them in no fixed order. What a twin can check across all of them is the student’s own understanding.

§04 Design step 3

When the two parts disagree

The threshold and the confirming weight from Step 4 both rest on what the authors call interdependence. One part has a say in what the other is worth, so a 50/50 split between a prepared presentation and an interview wouldn’t qualify. The framework assumes that a gap between the parts tells you something about the learning.

In their section on limitations the authors are more careful. A gap can have legitimate causes, such as anxiety, uneven skill development or less familiarity with one format, and two formats may not capture equivalent learning outcomes at all. They list reconciling conflicting grades among the barriers still to overcome.

Oral formats can be especially hard for students with social anxiety, other communication styles or another cultural background, and language barriers weigh on international students and others for whom English isn’t a first language. They still regard oral work as an important part of higher education, and they suggest supporting students with practice and scaffolding. Accommodations for disabled students have to cover both parts, they add, and Universal Design for Learning can help you anticipate diverse needs early.

Our addition. Decide in advance what a mismatch will lead to. We would treat it as a reason for a second conversation about the student’s drafts or notes before it becomes a mark. Nerves or language may explain the gap, so give credit for understanding even when a student expresses it haltingly or in a second language.

§05 Neighbours

Twins among other designs

The article presents twins as a way to triangulate evidence across formats that are good for learning and vulnerable to AI. Its examples of lower-risk modes for the second part are a supervised discussion, a quick derivation in class, an oral walkthrough and a process log.

Our addition. The process log is where twins meet Black Box Assessment (Winstone, Gravett and Elkington, 2026), which builds the process into the task and asks students to make their thinking visible through reflections, drafts and feedback encounters. Roe and colleagues add a second task that confirms the first. The two can be combined, and Black Box Assessment has its own deep dive.

§06 Read it for what it is

A proposal waiting for evidence

The article is conceptual, and no empirical data on the approach exist yet. The authors call the model a “bold and necessary experiment” (p. 10).

Twins cost staff time and need administrative and institutional support. Random sampling, the authors’ own answer for large cohorts, could weaken the validity the twin is meant to protect. For very large cohorts or tight resources they see a twin as possible but not optimal, and a complete redesign of the overall assessment strategy as more likely to help. Extra assessment also puts unequal pressure on students with paid work or family responsibilities.

Step 4 noted the risk the authors find perhaps most concerning, which is how twins are framed. Presented mainly as a way to catch AI misuse, they can build a surveillance mentality that treats students as potential cheaters and wears down the trust education depends on. Their conclusion points towards several assessment modes that work as complementary learning opportunities.

They call for pilot studies, mixed-methods research that includes student stress and staff workload, longitudinal studies of whether twins improve learning, and research into whether twin performance relates to competence outside the university.