Introduction:
The rapid emergence of generative artificial intelligence (AI) tools, including large multimodal models, has opened up new possibilities in dermatologic diagnostics, triage, and clinical education. However, evidence-based validation of these tools in real-world clinical workflows remains limited. This study aimed to evaluate the diagnostic accuracy, interpretability, and clinical usefulness of a generative AI-assisted dermatology system compared to standard clinician assessment.

Methods:
A cross-sectional, single-center study was conducted using 500 anonymized clinical images encompassing 15 common dermatologic conditions, including melanoma, basal cell carcinoma, acne, psoriasis, and eczema. A generative AI model (based on a large vision-language transformer) produced diagnostic suggestions and text-based clinical summaries. These outputs were compared against dermatologists’ diagnoses and histopathologic confirmations when available. Diagnostic accuracy, sensitivity, specificity, and time-to-decision were measured. Additionally, clinician surveys assessed the perceived utility, trust, and explainability of AI-generated outputs.

Results:
The AI model achieved an overall diagnostic accuracy of 83%, comparable to dermatology residents (84%) and superior to general practitioners (71%). Sensitivity for malignant lesions was 92%, though specificity remained moderate at 78%. Clinicians reported improved efficiency and patient education through visual-text generation but expressed caution regarding overreliance and potential bias in darker skin tones.

Conclusions:
Generative AI demonstrates strong potential to augment dermatologic care by improving diagnostic precision and patient engagement. However, rigorous validation across diverse skin types and transparent interpretability frameworks are essential before widespread clinical adoption. Continued evidence-based evaluation will ensure these technologies enhance, rather than replace, human expertise.