SYNTHESIS NOTE
Topics›Conversation Topics Dialog›this note

Can LLMs truly update shared conversational common ground?

Explores whether large language models can participate symmetrically in Stalnaker's picture of communication, where speakers mutually revise shared assumptions. The question matters because it reveals whether human-LLM dialogue is genuinely interactive or structurally asymmetrical.

Synthesis note · 2026-05-01 · sourced from Conversation Topics Dialog

On Stalnaker's picture, communication is a process of mutually proposing and accepting updates to shared assumptions. Each assertion is a candidate for incorporation into common ground; participants accept, query, or reject. The common ground evolves as conversation proceeds, and that evolution is itself the substance of communication.

LLMs cannot participate in this process symmetrically. The prompt establishes the model's working context, and the model interprets subsequent turns within that frame. Even when a user pivots — shifting from climate policy to historical precedent, or revealing they are not actually a five-year-old after asking for a five-year-old explanation — the LLM cannot smoothly absorb the revision into a jointly held common ground. It either ignores the pivot, fabricates continuity, or requires the user to re-scaffold from scratch. The asymmetry is structural: humans propose, the LLM either adopts or routes around, but the LLM cannot itself propose updates that change what counts as background.

This is a deeper deficit than failures of memory or inference. It means that the conversational scoreboard — Lewis's mechanism for tracking what counts as a felicitous next move — is one-sidedly maintained by the user. The user is keeping score for both players. The model is producing moves that look responsive but cannot reciprocally update the score in the way the conversational practice requires. What looks like dialogue is structurally closer to oracle-consultation, where the questioner provides all context and the oracle returns a response framed within it.

Inquiring lines that read this note 115

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

How can conversational AI develop appropriate initiative and goal awareness? Is mimicking human communication the best design approach for conversational AI? How do conversational patterns prevent dialogue failure and maintain understanding? Do face-saving social norms drive language model sycophancy and grounding failures? How do language models handle contested beliefs versus facts? Do language models presume shared understanding rather than build it? Why do LLM outputs diverge from their knowledge and reasoning? Do language models possess genuine mental states or only simulate understanding? Do language models learn genuine understanding or just surface patterns? How can AI best support therapeutic practice without replacing clinicians? What distinguishes genuine linguistic understanding from mere pattern matching? Do LLMs lack causal understanding needed for accurate prediction? When do complex LLM tasks benefit from external verification or structured scaffolding? How do multi-agent LLM systems fail at genuine coordination and deliberation? How do diffusion language models enable capabilities autoregressive architectures cannot achieve? How do conversational dynamics influence language model factual beliefs? Does RLHF training optimize perceived quality over practical accuracy? How can AI systems maintain consistent personas across extended dialogue? Do language models rely on encoded knowledge or primarily on surface patterns? Why do LLMs succeed at generation but struggle with evaluation? Does alignment training modify underlying model representations or just output compliance? How do social dynamics and peer behavior influence agent decision-making and cooperation? How do symbolic and semantic approaches combine in LLM reasoning?

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

Common ground in human-LLM conversation cannot be jointly updated because the LLM treats prompts as static frames