Key points are not available for this paper at this time.
Background and Objective: In recent years, there has been a dynamic development of artificial intelligence (AI), which has resulted in increased interest in its use in medical areas such as dietetics and psychodietetic. Current evidence remains insufficient to determine whether large language models (LLMs) are capable of generating diagnostically and therapeutically relevant responses in simulated psychodietetic scenarios. The aim of this study was to evaluate the performance LLMs—ChatGPT and Google Gemini—in simulated tasks related to diagnosis and psychodietetic intervention in anorexia nervosa (AN). Methods: Two complementary studies were conducted—the first, in which both models played the role of a patient suffering from anorexia and were tasked with answering questions asked in a psychodietetic interview, and the second, in which the chats analysed the case of a patient suffering from anorexia presented to them, and their task was to propose a correct diagnosis and therapy. Results: Both models have demonstrated the ability to engage with prompted scenarios and generate relevant, consistent responses. They were able to perform appropriate analyses, identify abnormal behaviors and formulate orderly diagnostic and therapeutic interpretations. At the same time, during the research, some differences were observed between the models in their interpretative approach and the level of development of the analyses. Conclusions: A key limitation of the models was a limited depth of empathic engagement and a reduced capacity for context-sensitive emotional responsiveness. Nevertheless, LLMs can potentially provide support in psychodietetic interventions if they are programmed correctly and are not abused in areas where they are not doing well enough.
Witkowska et al. (Wed,) studied this question.