Common ground

Common ground is the knowledge, beliefs and assumptions that two people mutually take to be shared, and grounding is the work they do in conversation to build and check it: introducing something, confirming it was understood, asking about it, repairing a misunderstanding. A speaker or writer who designs a message for a particular audience, what psycholinguists call audience design, has to estimate what that audience already knows; the usual error is to overestimate it, a bias known as the curse of knowledge. How far speakers actually track a partner’s knowledge is disputed: one line of work holds that people start from their own perspective and take the partner’s only as delayed repair, while others hold that speakers use simple partner cues that work often but not always (Horton & Brennan, 2016). The same problem now arises when an AI agent writes to a user in the middle of a task: studies of language models find that they produce far fewer grounding acts than people do (Shaikh et al., 2024), and the references, coined labels and status reports of an agent can depend on context the user no longer holds in mind.

History

The research did not cover the origin of the concept of common ground or the development of the grounding literature; add a history from a source that traces it.

Core concepts

Grounding and reference

In conversation, partners do not just transmit information; they work together to settle what is meant. When a hearer cannot identify what a speaker’s expression refers to, the two often collaborate on a new referring expression (Heeman & Hirst, 1995). A label proposed by one party is therefore not shared until the other accepts it or the pair rebuilds it. Grounding acts include restating one’s understanding in the form of a clarification question, as against assuming correctness and giving suggestions (Shaikh et al., 2024).

Conceptual pacts and shared labels

Partners who refer repeatedly to the same thing settle on a label. A pact on a label becomes stronger and more easily evoked with repeated reference (Horton & Brennan, 2016), and it is tied to the partner. In one study, after a pact had formed, addressees were slower to look at the object when the original partner switched to a new expression than when a new partner used it (Metzing & Brennan, 2003). With a new partner, speakers re-conceptualised the referent and spent substantial effort grounding their references again, as with the first partner in the first round; when the original partners were reunited, they returned smoothly to the earlier label, which the authors attribute to the partner’s presence cueing memory of the earlier episode (Horton & Brennan, 2016). The mechanism is itself debated: with episodes that differ only by partner, the distinction between memory traces can blur, so the partner effect need not reflect a partner-specific pact (Horton & Brennan, 2016).

Audience design

Audience design is the shaping of a message to its audience. Speakers who retold a story to a listener who had already heard it shortened it and left out detail, in events mentioned, number of words and amount of detail, compared with retellings to a new listener (Galati & Brennan, 2010). The authors argue that such audience design is computationally feasible with a “one-bit” model: has my audience heard this before, or not (Galati & Brennan, 2010). On this and the memory-based account, speakers do not keep full partner models. They design expressions from available representations that are likely but not guaranteed to work, and rely on the addressee to signal trouble (Horton & Brennan, 2016). How far common ground shows in speech depends on ordinary cognitive factors such as immediacy, relevance, perceptual vividness and goals, and on whether the information has already been processed in the conversation (Horton & Brennan, 2016).

The curse of knowledge

In a study of predictions about novices, experimentally reducing reliance on one’s own knowledge did not improve the accuracy of the estimates; the authors attribute the error to the lack of diagnostic cues about other people’s knowledge (Tullis & Feder, 2022). The same paper cites earlier work in which rewards and punishments for ignoring one’s own knowledge did not reduce egocentric bias in accounting decisions (Tullis & Feder, 2022). What did help was feedback on actual outcomes. In a three-condition experiment (no feedback, accuracy feedback, narrative feedback), feedback improved perspective judgments but only narrative feedback reduced the egocentric bias (Damen et al., 2021). In prediction tasks, the bias of valuing information that hurts the prediction of uninformed others shrank with experience and feedback on actual performance (Loewenstein et al., 2006).

Distance and accessibility of earlier content

A reference is easier to resolve when its antecedent is close. In adults, a sentence took less time to comprehend when the referents of its definite noun phrases had been mentioned one sentence back than two or three sentences back; later experiments placed the critical boundary at the clause rather than the sentence, which the authors read as the last clause processed giving its entities a privileged place in working memory (Clark & Sengul, 1979) (see working memory). A pronoun’s boost to its referent faded by the end of the next sentence, though it lasted longer in longer, richer stories, which the authors attribute to engagement rather than distance (Love & McKoon, 2011).

Repetition and expertise

Whether repeating something the reader knows is a cost depends on the reader. A review of 63 multimedia-learning studies finds that repeating the same content hinders learners with high prior knowledge while novices may benefit; information useful to novices becomes redundant for more knowledgeable learners (Trypke et al., 2023). The review also reports that the benefit of adding text to narration depends on prior knowledge, and its authors call for a more consistent theory of redundancy (Trypke et al., 2023).

Resuming after a break

After an interruption, the common ground between a person and their own earlier work has to be rebuilt. The goal-activation account holds that goals decay, and that rehearsal and environmental cues are the two ways to slow the decay (Trafton et al., 2005). A blatant cue marking the last action (a red arrow) shortened resumption time, while a subtle cue (the cursor at the last action) was no better than no cue (Trafton et al., 2005). Time to encode retrieval cues before leaving the task reduced the cost of resuming, but the benefit was impeded when the visual cues changed during the interruption (Hodgetts & Jones, 2006). In a lab study with programmers, automated cues showing the user’s prior activity doubled task success after an interruption compared with the programmers’ own notes (Parnin & DeLine, 2010). In reading studies, restating the last one or two sentences before an interruption countered its cost, while restating thematic information did not (Glanzer and colleagues, as reported by Ledoux and Gordon) (Ledoux & Gordon, 2006).

Structured handovers

Clinical handovers are a worked example of deliberately building common ground between people who do not share a context. The I-PASS format has five elements: illness severity, patient summary, action list, situation awareness and contingency planning, and synthesis by the receiver (Starmer et al., 2022). The SBAR family of formats separates the current state (Situation) from the history (Background) (Lazzari & Rabottini, 2025). See the controversies below for the strength of the evidence.

Questions, controversies, and new developments

Do speakers model their partners, or start egocentrically?

One line of work holds that speakers and addressees are inevitably egocentric, and that taking account of a partner’s perspective can happen only as delayed processing or repair (Horton & Brennan, 2016). The one-bit and memory-based accounts instead hold that partner information is used when it is available and relevant, with failures following from ordinary factors such as immediacy and vividness (Horton & Brennan, 2016; Galati & Brennan, 2010). The question is not settled.

Presumptive grounding in language models

Off-the-shelf language-model generations were on average 77.5 percent less likely to contain grounding acts than human turns given the same conversational context (Shaikh et al., 2024). Supervised fine-tuning did not improve grounding and preference optimisation eroded it (Shaikh et al., 2024). A second study of human-LLM interaction logs found that LLMs were three times less likely to initiate clarification and sixteen times less likely to make follow-up requests than humans, and that early grounding failures cascaded into higher likelihoods of later failure (Shaikh et al., 2025). Its authors propose a forecaster of when a grounding act is needed as an intervention, but say that progress will require rethinking how models are trained to interact with people (Shaikh et al., 2025). The two papers have overlapping authors.

Asking before answering can help, on thin evidence. In one user study, documents written after follow-up questions were preferred, and users valued open-ended or thought-provoking questions over simple information-gathering ones (Tix, 2024). Allowing clarification on only 10 percent of examples, chosen by an uncertainty estimate, doubled the gain of choosing the examples at random (Zhang & Choi, 2023). A small rubric comparison found that a clarifying-question prompt scored 6.67 of 8, against 5.67 for a raw prompt and 7.50 for a checklist-improved prompt, and used more tokens than the checklist prompt (Ghosh, 2026).

Benchmarks of human-AI grounding

A collaborative puzzle benchmark measures situation awareness, learning effects, communication efficiency and grounding behaviours (Poelitz et al., 2026). When the person could not see the AI’s actions, the AI often failed to clarify or repair and failed to update its assumptions after a repair, and both sides reduced their grounding effort instead of increasing it, as work on human pairs would predict (Poelitz et al., 2026). People adapted to the AI’s identifier labels and carried the load of remembering them across trials unless they asked the AI to list them, which the authors read as a convention pact (Poelitz et al., 2026). The measures were message and word counts, exact match of the final puzzle to a target, and analysis of reference types and LLM-labelled dialogue acts (presentation, clarification, repair, acceptance) (Poelitz et al., 2026). In a repeated-round picture-matching task, vision-language models could not generate and resolve referring expressions interactively well enough for smooth communication (Zeng et al., 2026). Both are preprints with one task each.

Losing the thread over many turns

In simulations where task details arrived over several turns, all top models tested scored lower than when given the details in one turn, with an average drop of 39 percent across six generation tasks, and they did not recover from early wrong assumptions (Laban et al., 2025). The strongest models were as affected as smaller ones (Laban et al., 2025). Repeating all earlier user information on each turn recovered 15 to 20 percent of the loss; a single recap at the end did better but is unrealistic, since the last turn is not known in advance (Laban et al., 2025). In a separate test, agents given a goal in the system prompt drifted toward competing objectives over long contexts in every model tested, with drift correlating with growing pattern-matching as context grew (Arike et al., 2025). These are simulated users and older models, so they show a mechanism, not current rates.

Observed failures of coding agents

In an observation of 19 developers using an in-IDE agent, several found its explanations too verbose, and verbosity once caused a participant to miss a link in the response; participants wanted to see the changes rather than long analysis (Kumar et al., 2025) (see verbosity in language models). When the agent changed more than the prompt asked, a participant had to reject the change and write a follow-up prompt restricting it (Kumar et al., 2025). Interviewed engineers expected agents to explain the reasoning behind suggestions, to surface knowledge gaps proactively and to ask questions, and described the agent as a black box; users’ own rules include limiting actions to the request and presenting a plan for approval before execution (Dong et al., 2025). The authors of a goal-tracking tool note that in long chats it becomes hard to trace which requests were met or contradicted earlier ones as users scroll an ever-growing log, and in their 20-person study many baseline participants scrolled through lengthy logs to find what went wrong (Coscia et al., 2025). The same paper cites earlier work that chats in writing and information-foraging tasks often stay at 10 turns or fewer because of mental fatigue from the growing log (Coscia et al., 2025).

A persistent graph of the project beside the chat was tested against a chat-only baseline in a randomised between-subjects study of 16 developers; it supported project comprehension and task performance (Chou et al., 2026). Its authors state that delegating implementation can build up cognitive debt that erodes project comprehension over time and limits oversight (Chou et al., 2026). This is a preprint, and it tests no message-wording rules. A spoken-discourse poster found that interference from distance to the antecedent appeared in monologue but speaker cues reduced it in dialogue; it concerns ellipsis in speech, not written chat (Martin, 2016).

What status reports should contain

A framework of agent transparency defines three levels: what the agent is doing, why, and what it will do next (Daronnat et al., 2022). In an experiment with 180 participants, explanations of the first level (what the agent is doing) led to better task performance and the second (why) increased trust; the paper’s discussion contradicts its abstract on descriptive versus prescriptive explanations (Daronnat et al., 2022). For handovers, I-PASS adoption was associated with a rise in written handoffs containing all five elements from 10.4 to 73.5 percent and a 47 percent fall in handoff-related reported adverse events (Starmer et al., 2022), but the review of the SBAR-family formats notes a lack of randomised trials and mostly cross-sectional designs, so its conclusion that they improved quality of care is weak evidence (Lazzari & Rabottini, 2025). The resumption evidence is mixed too: earlier work found that a visible marker did not help resumption, which Trafton and colleagues attribute to coarse measures or markers that were interface actions (Trafton et al., 2005).

Broader connections

Common ground links the psychology of language to the practical problem of writing for a reader whose current knowledge differs from the writer’s, taken up in writing for comprehension and explanation. Accessibility of recent content depends on working memory. The handover and status-report formats above parallel the relation of a delegator to a delegate, as in the manager-employee relationship.

Further reading

• Horton and Brennan (2016): a review of egocentric, one-bit and memory-based accounts of how speakers and addressees use partner information.
• Galati and Brennan (2010): the retelling experiment behind the one-bit model of audience design.
• Tullis and Feder (2022): the curse of knowledge in predictions about novices, and what failed to reduce it.
• Shaikh et al. (2024): the measurement of missing grounding acts in language-model output.
• Laban et al. (2025): how models degrade when task details arrive across turns.
• Poelitz et al. (2026): a benchmark of grounding in human-AI puzzle collaboration.