Accepted · ACDSA 2027
EVALUATION / Accepted · ACDSA 2027
An explanation for every audience?
Examining readability, safety, and human–AI agreement in audience-targeted explanations of brain MRI images.
01 / MOTIVATION
The question
Can audience-specific prompting make an explanation more accessible while preserving reliable image interpretation?
02 / METHOD
The approach
Four audience-specific prompts are evaluated on 400 brain MRI images spanning four classes, yielding 1,594 valid explanations.
Evaluation covers readability, hallucination, safety, and audience fit. Human and AI ratings are compared on 160 shared responses, and explicit tumor-class mentions are checked against dataset labels.
03 / EVIDENCE
Findings & scope
Patient and student prompts produced simpler explanations. Exact human–AI agreement on hallucination ratings was only 4%.
Audience adaptation did not ensure reliable class identification. This is an evaluation of generated explanations, not evidence of clinical diagnostic readiness.
READ THE WORK