EVALUATING THE ACCURACY OF GENERATIVE AI IN PRODUCING CULTURALLY RELEVANT RECIPE RECOMMENDATIONS FOR HEALTH APPLICATIONS

Culturally relevant meal recommendations have been shown to aid in long-term diet adherence and positive attitudes towards healthy eating. However, as interest in AI-driven food and health recommendations grows, the cultural accuracy of popular large language models (LLMs) is unknown. This study evaluates the accuracy of culturally relevant recipes generated by Google’s Gemini LLM by comparing them to real recipes obtained from a culturally diverse, reputable online resource, AllRecipes.com. We applied various lexical and semantic feature extraction methods in combination with diverse classification methods to determine which pairing is most appropriate when looking at recipe data. Our findings align with previous research that simpler methods, like word frequency-based embeddings and Naïve Bayes classification, yield the highest accuracy in classifying a recipe’s cultural origin. We further improved the usability of our classifier by implementing hierarchical clustering to group culturally similar cuisines and reduce classification complexity. Overall, our findings demonstrate the viability of using generative AI in culturally sensitive, health-conscious recipe recommendations and add to the larger body of work verifying generative AI output to ensure human safety.