LLM-based healthiness? Analyzing the nutritional quality of an AI-generated recipe dataset

Recipe discovery has become increasingly accessible through large language models (LLMs). However, the nutritional content and quality of AI-generated content remain underexplored, raising concerns about its public health implications and its applicability in food recommender systems. This paper presents a nutritional evaluation of a recipe dataset generated by an LLM, assessed against health guidelines from the World Health Organization (WHO) and the UK Food Standards Agency (FSA). Using standardized scoring systems, we analyzed subsets of general and diet-specific recipes. Although the model produced moderately healthy recipes, it often did not meet stricter dietary standards, particularly in terms of fiber content and saturated fat levels. Nutritional quality also varied between meal categories, with desserts and snacks performing the worst. In addition to these findings, we introduce a curated and labeled dataset of recipes generated by an LLM to support future work in health-based food recommendation. Our results highlight both the potential and limitations of generative models in nutrition-sensitive domains.