Large language model responses to structured textual descriptions of salivary gland imaging findings: impact of contextual enrichment on self-reported diagnostic confidence and reporting behavior.

PURPOSE: Large language models (LLMs) are increasingly applied in clinical decision support, yet their behavior when presented with structured textual descriptions of imaging findings, particularly in salivary gland disease, remains poorly defined. The present study was set out to evaluate how the responses of different LLMs change as the textual input is progressively enriched. The models did not receive imaging data; all inputs were text. METHODS: This exploratory, simulation-based observational study assessed 3 LLMs (ChatGPT, Gemini, Claude) across 10 salivary gland cases. For each case, a standardized textual description of the imaging findings was prepared and submitted as a text prompt; no image files were uploaded to any platform. Each case was evaluated under 3 input conditions of increasing informational content: single-plane textual description, multiplanar textual description, and multiplanar textual description plus clinical information, generating 90 total observations derived from repeated measurement of the same 10 cases. Primary outcomes were the model-generated severity score, self-reported diagnostic confidence, and the number of additional reported anatomical structures. AI outputs were not benchmarked against a verified reference diagnosis. The primary analysis was a proportional-odds model with input condition as the fixed effect of interest, AI system as a covariate, and cluster-robust standard errors by case; non-parametric tests were used as sensitivity analyses. RESULTS: Model-generated severity scores did not significantly differ among LLMs (χ²=2.31; P = 0.31). Self-reported diagnostic confidence increased significantly as the textual input was enriched (Friedman χ²=18.86; P < 0.001; proportional-odds model, multiplanar description plus clinical information vs. single-plane description, cumulative OR 5.43, 95% CI 3.00-9.81). ChatGPT produced stable outputs with no additional reported structures, whereas Claude produced substantially more detailed reports (mean additional structures, 5.67 vs. 0.00). Gemini produced lower severity scores but comparable self-reported confidence. Mixed-model analysis revealed a significant effect of AI system on reporting verbosity (P < 0.001). Associations between model-reported parenchymal changes and self-reported confidence were also observed; because both variables were generated by the same model within the same response, these represent within-output associations rather than independent diagnostic relationships and are reported descriptively. CONCLUSION: In this exploratory pilot study, LLMs produced comparable severity ratings, while self-reported diagnostic confidence increased as more information was supplied in the prompt. This pattern may reflect the tendency of an LLM to express greater certainty when given richer input rather than genuinely improved diagnostic reasoning. Substantial variability in reporting behavior exists across models, particularly in output detail. These findings suggest that LLMs may warrant further investigation as adjunctive tools for text-based tasks in salivary gland care. The present design does not evaluate image interpretation and does not support use as standalone diagnostic systems.