How do prompts reshape the role of context in AI conversation?

In human dialogue, context is partly inherited as common ground and partly built incrementally through cooperative conversational moves, with each speaker adjusting framing based on real-time feedback from the other. With an LLM, the user must scaffold context unilaterally through a single prompt — describing intended audience, register, role, and topic in advance. This makes the prompt a categorically novel speech act: simultaneously utterance, common-ground assignment, role allocation, and goal specification compressed into a frame the LLM treats as static.

Kasirzadeh and Gabriel compare this to a theatre director setting stage, lighting, and script in advance before a performance — the actor must perform within those specifications rather than negotiate them. Two consequences follow. First, priming becomes explicit and exhaustive rather than backgrounded and dispositional, contradicting the implicit-knowledge view of context that runs from Searle's Background through ordinary-language philosophy: the LLM cannot use the kind of unconscious practical know-how that lets a hearer of "cut the cake" reach for a knife rather than a lawnmower. Second, the conversation cannot evolve beyond what the prompt anticipates; mid-conversation pivots require explicit re-scaffolding, or the LLM defaults to the original frame.

This formalizes what Language as Event names directly. The LLM does not produce utterances inside a shared event. It produces residue that the human must convert into a pseudo-event by supplying the orientation unilaterally — and the prompt is the site where that asymmetric labor is paid.

Source: Conversation Topics Dialog