Describe or Direct? Artificial Intelligence Explanation Type, Trust and User Behaviour in a Medical Imaging Rostering Task.

INTRODUCTION

Artificial intelligence (AI) is increasingly integrated into healthcare, with most healthcare AI applications concentrated in radiology. While explainable AI (XAI) has been proposed as a means of improving transparency and trust, the relationship between explanation features, trust and user behaviour remains underexplored, particularly in healthcare contexts. This study examined how final-year medical imaging students interacted with descriptive and prescriptive AI explanations during a simulated rostering task, with a focus on trust and user behaviour.

METHODS

A human-grounded evaluation using an explanatory mixed methods design was undertaken. Final year medical imaging students completed simulated rostering tasks using an XAI system that generated either descriptive or prescriptive explanations. Quantitative measures collected included task completion metrics, explanation regenerations and self-reported trust scores. Participants took part in semi-structured discussion groups to explore user perceptions, explanation preferences and trust dynamics.

RESULTS

Trust scores increased across the experiment for both explanation types. Prescriptive explanations were associated with significantly fewer explanation regenerations and higher trust scores. Task time and failed solutions showed small-to-moderate effects favouring prescriptive explanations, but these differences did not reach statistical significance. Qualitative findings indicated a strong participant preference for prescriptive explanations, alongside critical engagement with inaccuracies and regeneration patterns.

CONCLUSION

Prescriptive AI explanations were preferred over descriptive explanations within a simulated medical imaging rostering task and were associated with higher trust and reduced explanation regeneration. Participants demonstrated critical engagement with AI-generated outputs rather than uncritical acceptance. These findings contribute to emerging evidence regarding human interaction with XAI systems in healthcare workflow settings.