(This talk will take place in LORIA, room A008.) LLMs are used in conversational agents (CAs) (e.g., Claude, ChatGPT) to interact with users through text generation. The sheer number of parameters in LLMs make it difficult to explain and predict the behavior of these CAs. Several techniques have been developed for this purpose: e.g., activated software neurons visualization, seeking “human functions” amongst them, and even explicitly asking for generated explanations by the CA itself. We propose the adoption of the radical behaviorist’s stance described by Skinner in the design and study of experiments that seek to understand and predict the verbal behavior of CAs. We explore the epistemological feasibility of such a stance in the study of CAs, specifically by the study of their intraverbal behavior through a free association experiment.