Making marking visible: New evidence for AI in assessment

Using eye-tracking to inform an AI virtual examiner assistant

Published

Friday 6 Nov 2026

Close-up image on a human eye with a square target superimposed over it

How do experienced examiners work towards a marking decision? Can evidence about
that process help inform the design of AI-enabled examiner support? Our researchers
used eye-tracking and multimodal process data to investigate the cognitive work
behind marking decisions.

The aim: to build a stronger evidence base for AI-enabled examiner support that reflects actual practice.

Why this matters

Marking involves more than applying a mark scheme to a student response. Examiners interpret, compare and synthesise information from different sources as they work towards a judgement. Much of this activity is not visible in traditional marking data, but can be revealed through patterns of attention, navigation and
interaction.

What the research did

The study analysed how 25 experienced examiners approached marking across 750 tasks, using a combination of eye-tracking, log data, think-aloud evidence and self-report measures. Together, these methods provide detailed insight into how examiners interact with marking materials as they work towards a judgement.

A new way of understanding marking

Eye-tracking data enabled the researchers to identify different patterns of activity during marking, from initial scanning and orientation to focused reading, integration of information across marking materials, evaluation of the student response and movement towards a final decision.

Eye-gaze attention is distributed across multiple sources – not fixed on the response.

Eye-gaze attention is distributed across multiple sources – not fixed on the response.

This provides a way of examining marking as a dynamic process rather than treating the final mark as the only observable outcome.

Key insights

  • Marking is a process not a moment
    Examiners move between integrating and evaluating information as they work towards a judgement.
  • There isn’t one ‘correct’ way to mark
    Different examiners follow different pathways – even when they agree on the mark.
  • Marking relies on more than the student response
    Examiners spend significant time using mark schemes and commentary alongside the response to build and check their judgement.
  • How marking materials are presented matters
    The structure and presentation of the marking materials influences what examiners attend to and how information is integrated during marking.
  • Examiners use marking resources differently
    Individual variation in how examiners navigate and integrate material suggests that effective support needs to respond to the examiner’s context and stage in the process.

What this could mean for future AI-enabled support

Bringing information together

AI-enabled support may need to help examiners bring together information from multiple sources, reflecting how examiners build judgements using mark schemes, commentary and other guidance.

Respecting examiner thinking time
Effective support should respect examiner thinking time, providing assistance when required and desired without disrupting the judgement process.

Adapting to different approaches
Different examiners may follow different pathways while arriving at similar judgements, raising questions about how AI-enabled support can adapt to different marking approaches.

Learning from process data

Process data may provide new opportunities to understand examiner behaviour and inform the development and evaluation of future support tools.

Further research is needed to explore which approaches are most effective in practice.