Abstraction-dependent diagnostic performance of a multimodal foundation model in oral epithelial dysplasia.

The reliability of general-purpose multimodal large language models (LLMs) in oral histopathologic image interpretation remains incompletely characterized. To evaluate abstraction-level diagnostic competence, grading reliability, and structural error patterns of ChatGPT-5.2 in oral epithelial dysplasia histopathology. In this retrospective diagnostic accuracy study, ChatGPT-5.2 analyzed 200 digitized H&E-stained oral mucosal images (100 OPMD; 100 normal) in a zero-shot setting. Model outputs were compared with consensus diagnoses from three expert oral pathologists. The primary outcome was abstraction-level agreement (κ) across predefined WHO-aligned morphologic domains. Secondary outcomes included binary diagnostic accuracy, grade-stratified sensitivity, grading discordance, feature-level error profiling, inter-run stability, and modeled clinical utility. Agreement declined monotonically across morphologic abstraction domains (κ: 0.85 coarse morphology; 0.42 architectural; 0.09 high-risk cytologic; P[ordered] ≈ 1.000). Binary classification achieved a sensitivity 87.0% (95% CI 79.0-92.6) and a specificity 89.0% (95% CI 81.4-94.0). All severe dysplasia cases were detected (100%), with false negatives confined to non-severe lesions. Grading agreement was fair (weighted κ = 0.36) with predominant under-grading. Feature-level degradation was omission-dominant and concentrated within high-risk cytologic descriptors. Binary outputs demonstrated high inter-run stability (Fleiss' κ = 0.82). ChatGPT-5.2 demonstrated stable binary discrimination for OPMD detection but showed abstraction-dependent degradation in architectural and cytologic feature recognition central to dysplasia grading. While performance may support assistive or triage-oriented applications, grading variability and omission of high-risk cytologic criteria indicate that expert oversight remains essential. Further domain-specific validation is required before clinical integration.