Large Language Models (LLMs) are often described as ‘conversational’; they are often termed as conversational AI, and interactions with them are frequently called conversations. In this dissertation, I depart from this mainstream view and argue that LLMs should not be seen as capable of participating in conversation. Utilising the view that conversation is rooted in negotiated mutual social commitments, I firstly argue that LLMs cannot be seen as socially committed to the truth of their outputs.…
Read moreLarge Language Models (LLMs) are often described as ‘conversational’; they are often termed as conversational AI, and interactions with them are frequently called conversations. In this dissertation, I depart from this mainstream view and argue that LLMs should not be seen as capable of participating in conversation. Utilising the view that conversation is rooted in negotiated mutual social commitments, I firstly argue that LLMs cannot be seen as socially committed to the truth of their outputs. As such, I contend, they are unable to participate in conversation. The argument then turns to the mutuality of commitments in conversation. I argue for an interpretation of LLM outputs where they each output is generated by a different interlocutor after each prompt. The conclusion reached is that since there is no consistent LLM interlocutor, the commitments can never be mutually held, which would entail that we cannot have a conversation with an LLM. As such, I have aimed to argue that using the term 'conversational' to describe LLMs is a misnomer.
