CAN CHATGPT PROVIDE RELIABLE DIETARY ADVICE? EVALUATION OF LLMS VIA CHAIN-OF-THOUGHT AND NUTRITIONAL LIMITATIONS IN THE BRAZILIAN CONTEXT

Abstract

The use of Large Language Models (LLMs) for dietary prescription raises clinical safety concerns. This study compares the efficacy of ChatGPT, Gemini, and DeepSeek in generating hypocaloric plans for overweight Brazilian women. Using one-shot and chain-of-thought prompts, 150 meal plans were generated and nutritionally analyzed via TBCA. Results show energy variability: ChatGPT presented superior precision (<0.5% deviation), while the other models overestimated calories. Qualitatively, all models failed: saturated fat exceeded limits by over 200% (reaching 291% in DeepSeek), and systematic micronutrient inadequacies were observed, with iron reaching only 55% to 61% of the recommendations, alongside a calcium deficit. It is concluded that the qualitative inaccuracy of LLMs prevents their autonomous clinical use, requiring professional supervision.