Alignment of large language model–generated meal plans with national nutrient reference values: A comparative analysis of five models

BackgroundLarge language models (LLMs) increasingly generate dietary advice, but their alignment with country-specific nutritional guidance remains uncertain.AimTo descriptively evaluate the alignment of the first seven-day meal plans generated by five publicly accessible LLMs with the Türkiye Dietary Guidelines (TÜBER) 2022 reference values for a 2000-kcal diet.MethodsGPT-5.3 (ChatGPT), Gemini 2.5 Flash, Grok 4.2, the default free-web Microsoft Copilot model, and DeepSeek-V3 received an identical Turkish-language prompt describing a standardized healthy adult. Each model's first seven-day plan was analysed using nutrition analysis software. Percentage deviations from TÜBER references, Dietary Diversity Score and Food Variety Score were calculated and evaluated descriptively.SummaryAll plans had high dietary diversity. Mean energy ranged from 1378 to 1696 kcal/day, below the 2002-kcal reference. Fibre ranged from 23.3 to 33.5 g/day versus the 35-g reference, and riboflavin ranged from 1.43 to 2.19 mg/day versus 2.4 mg. Sodium ranged from 1759 to 2938 mg/day, exceeding the 1549-mg reference, whereas calcium ranged from 712 to 1032 mg/day versus 1284 mg. Deviations varied across plans, and no model was consistently closest to the references across all nutritional components. High dietary diversity did not ensure alignment with national nutrient reference values. LLM-generated meal plans require verification against national guidance and professional review before individualized use.