Daniele Roberto Giacobbe,1,2 Alessandra Agnese Grossi,3,4 Cristina Marelli,5,6 Marco Muccio,1 Sabrina Guastavino,7 Ylenia Murgia,8 Sara Mora,9 Alessio Signori,10,11 Nicola Rosso,9 Mauro Giacomini,8 Cristina Campi,7,12 Michele Piana,7,12 Matteo Bassetti1,2 1UO Clinica Malattie Infettive, IRCCS Azienda Ospedaliera Metropolitana, Genoa, Italy; 2Department of Health Sciences (DISSAL), University of Genoa, Genoa, Italy; 3Department of Human Sciences and Innovation for the Territory, University of Insubria, Varese, Italy; 4Department of Biotechnologies and Life Sciences, Center for Clinical Ethics, University of Insubria, Varese, Italy; 5CESP - INSERM U1018, Oncostat, Labeled Ligue Contre le Cancer, Gustave Roussy, Université Paris-Saclay, Villejuif, France; 6Institut Curie - INSERM U1331, Team Statistics Applied to Personalized Medicine, Paris, France; 7Department of Mathematics (DIMA), University of Genoa, Genoa, Italy; 8Department of Informatics, Bioengineering, Robotics and System Engineering (DIBRIS), University of Genoa, Genoa, Italy; 9UO Information and Communication Technologies, IRCCS Azienda Ospedaliera Metropolitana, Genoa, Italy; 10Department of Health Sciences (DISSAL), Section of Biostatistics, University of Genoa, Genoa, Italy; 11Ospedale Policlinico San Martino, IRCCS Azienda Ospedaliera Metropolitana, Genoa, Italy; 12Life Science Computational Laboratory (LISCOMP), IRCCS Azienda Ospedaliera Metropolitana, Genoa, ItalyCorrespondence: Daniele Roberto Giacobbe, Department of Health Sciences (DISSAL), University of Genoa, Via A. Pastore 1, Genoa, 16132, Italy, Email danieleroberto.giacobbe@unige.itAbstract: General-purpose large language model (LLM)-based chatbots are increasingly used by clinicians to discuss medical problems, including antibiotic prescribing. Their use creates an unprecedented setting for clinical reasoning in which diagnostic and therapeutic thinking becomes dynamically shared between human and machine. Here, we propose a theoretical framework, intended for subsequent empirical assessment, around the concept of shared reasoning fragility, defined as the potential instability arising from the interaction between clinician reasoning and the chatbotâs opaque, association-based processes, which are structurally different from classical human reasoning. The theoretical framework is based on the conceptual argument that, while the black box dilemma is often discussed for classification-oriented clinical decision support systems with an emphasis on explainability versus external validation, chatbot-assisted practice introduces a distinct problem: chatbots can accompany clinicians throughout the entire reasoning pathway rather than being consulted only at the final decision point. In the present perspective, we argue more explicitly that the fragility of this continuous co-reasoning primarily stems from its novelty and pervasiveness. Using strictly illustrative examples in antibiotic prescribing, we suggest the theoretical possibility that fluent and convincing outputs may redirect attention, mask omissions in work-up, and subtly shift hypothesis selection during shared clinical reasoning processes. While it is important to stress that our framework is purely theoretical and thus cannot be confirmed at the present stage, our considerations are intended to motivate the required quantitative research to confirm or refute shared reasoning fragility, measure its extent, and evaluate downstream implications for patient care.Keywords: healthcare, infection, antibiotic prescribing, artificial intelligence, machine learning, deep learning, natural language processing
Giacobbe et al. (Fri,) studied this question.