Abstract
Background/Objectives: Patients with multimorbidity often receive conflicting dietary guideline recommendations, and large language models (LLMs) increasingly draft dietary advice. In this study, we evaluated nine LLMs, including two Turkish-adapted models, on 76 synthetic patient vignettes in Turkish and English. Methods: The vignettes combined type 2 diabetes, chronic kidney disease, hypertension, obesity, gout and celiac disease as single conditions, conflicting pairs, concordant pairs and triples. An LLM judge panel graded the responses against key-feature checklists approved by a dietitian, whose blinded ratings of a subset agreed with the panel on 90% of items. Results: The proprietary models covered 81% to 92% of the weighted checklist and fully resolved 57% to 89% of the guideline conflicts. Their coverage did not differ significantly between Turkish and English in any of these models, whereas across all models Turkish responses covered 2 percentage points less, mainly in conflicting-pair vignettes and in the open-weight models. The GPT and Claude models had the fewest safety-critical flags, 5% to 8% in both languages, and the proprietary models’ Turkish advice was rated 4.9 to 5.0 out of 5 for cultural appropriateness. Conclusions: Expert review of LLM drafts can therefore focus on conflict resolution, and the checklist procedure provides a reusable audit.