Use of Artificial Intelligence in obtaining canine diets: nutritional inadequacy and the need for technical knowledge

This study evaluated the nutritional adequacy of dog diets formulated by four widely used Artificial Intelligence models (Manus, ChatGPT, DeepSeek, and Gemini) against the recommendations of the European Pet Food Industry Federation. This was a theoretical, in silico study, no animals were fed or clinically evaluated. Using a standardized prompt, complete and balanced diets (beef and chicken) for an 8 kg adult dog were requested. Eleven formulations were analyzed using SuperCracPet® software based on United States Department of Agriculture data. Overall, 61.3% of nutrients were below the minimum recommended levels. Nutrient levels below FEDIAF minimum requirements were identified across multiple categories: energy supply was insufficient in 81% of diets, calcium levels were below recommendations in 91% of formulations, calcium-to-phosphorus ratios were outside the recommended range in 63% of diets, chlorine was absent in all formulations, vitamin D levels were below requirements in 63.7% of diets, choline in 81.8%, and linoleic acid in 72%. Only 18% of the formulations included supplementation, which was provided in insufficient quantity. If implemented chronically, these nutritional inadequacies may pose risks to bone, dermatological, immune, and metabolic health. It is concluded that, in their current form, the evaluated language models are insufficient for the autonomous formulation of canine diets, as the generated prescriptions presented multiple nutritional inadequacies. Safe application requires hybrid systems with mandatory supervision by animal scientists and veterinarians, as well as regulation to prevent avoidable risks to animal health.