Co‐textual dopes: How LLMs produce contextually appropriate text in chat interactions with humans without access to context

This paper asks how LLM‐based systems can produce text that is taken as contextually appropriate by humans without having seen text in its broader context. To understand how this is possible, context and co‐text have to be distinguished. Co‐text is input to LLMs during training and at inference as well as the primary resource of sense‐making for humans in interaction, collaboratively produced by both human and machine during chat. Systems can also passively participate in contextualization, insofar as cues are found in the co‐text and the user guides the system as to what is the appropriate context for them.