Is common ground the same as a knowledge base?

No. A knowledge base stores information; common ground concerns what participants treat as shared in an interaction.

Must it be stated explicitly?

No. Context and prior interaction can support shared assumptions, though important uncertain details may need clarification.

Does grounding here mean retrieving evidence from documents?

Not specifically. This conversational sense concerns establishing understanding between participants; other AI uses of grounding have different scopes.

AI Glossary / Natural language and speech / Intermediate

Common ground

Definition

Common ground is the information that participants in a conversation treat as shared, including relevant knowledge, beliefs and assumptions. It gives them a basis for interpreting references, coordinating actions and deciding what needs to be explained.

What it means

If two colleagues have already identified a particular presentation, one can say “move the last slide” without naming the file again. The phrase works because the conversation supplies a shared reference. If they are actually viewing different files, the apparent common ground is faulty.

Common ground is more than two people independently possessing the same fact. They also proceed on the assumption that it is shared for the current interaction. It can include information established during the exchange, the visible situation and background assumptions relevant to the participants.

The concept is not restricted to friendly agreement. People can disagree about a proposal while sharing an understanding of what the proposal is. Nor does information become objectively true merely because participants take it for granted.

How it works

Participants establish and update common ground through grounding: providing evidence that an utterance has been understood well enough for the current purpose. A relevant reply, a clarification, a paraphrase or an acknowledgment can contribute that evidence.

The required precision depends on the task. Casual conversation can proceed with a broad understanding; arranging an appointment may require a date, time and location to be made explicit. When a reference fails, a repair such as “I meant the revised file” can correct the shared working interpretation.

For an AI assistant, a dialogue-state record can represent confirmed details, unresolved questions and candidate interpretations. That record is a practical model of the interaction, not proof that the system has human-like mutual beliefs. The assistant should expose consequential assumptions in its wording and incorporate corrections instead of treating every earlier guess as established fact.

Three practical examples

A scheduling assistant confirms which of two proposed dates the user selected before creating an appointment.
A collaborative editing tool makes the active document visible so “this paragraph” has a shared referent.
A support agent restates a reported problem and lets the customer correct an incorrect interpretation before troubleshooting continues.

Think of it this way

Think of a shared working map with landmarks that both travelers believe they have identified. Directions can be shorter when the same landmarks are understood. If one person means a different bridge, they must repair the map before continuing.

The analogy concerns coordinated interpretation. Common ground is not a literal file containing everything both people know, and participants can mistakenly think their maps agree.

Worked example

Repairing an ambiguous reference

Consider an invented exchange. User: “Move the review to Friday.” Assistant: “The product review or the budget review?” User: “The budget review, at 10 a.m.”

At this point, the meeting type and intended hour have been supplied, but the exact Friday, time zone and authorization to change the meeting may still depend on the established context. A useful response can state the resolved details: “The budget review on Friday 16 October at 10 a.m. Dubai time?” The user confirms or corrects them.

This confirmation creates evidence of a shared interpretation of those details. It does not prove that a calendar update has succeeded. A later success report should be based on the calendar operation itself.

The example separates three layers: interpreting the request, establishing the relevant shared details and carrying out the change. Collapsing them into one generic “OK” can hide a mistake.

Illustrative case study

A fictional equipment-rental service deploys a conversational booking assistant. Customers often say “the same one again,” referring to different previous rentals. The assistant initially selects the most recent item without showing its interpretation.

The team changes the conversation to name the candidate item and relevant booking details. If the customer corrects the choice, the assistant updates its state and uses the corrected item in subsequent replies. Reviewers test whether later turns retain that correction.

The service does not require confirmation of every harmless detail. It concentrates on references that determine the booking. The lesson is that a shorter conversation is not necessarily more efficient if hidden misunderstandings later require rework.

Common misconceptions and mistakes

Common ground means identical opinions. Shared understanding can coexist with disagreement.
Everything in the conversation history is confirmed. Questions, guesses and rejected proposals have different status.
An acknowledgment proves complete understanding. Its meaning depends on context and the task.
Shared assumptions are always true. Participants can share a mistaken belief.
Giving the model a long context window automatically establishes common ground. Representation and interaction still need to track what has actually been settled.

Business use

Common-ground design improves handoffs, booking flows and collaborative work by making important assumptions visible. Identify which details control the next action and whether those details have been supplied, inferred or confirmed. Preserve that distinction in both summaries and internal records.

Evaluate conversations with corrections, multiple possible referents and changes of plan. A useful measure is whether the assistant completes the intended task after clarification, not merely whether its replies sound natural. When another human or system takes over, communicate unresolved points as clearly as confirmed information.

Check your understanding

Why can two people know the same meeting time but still fail to coordinate?

Answer: They may not know that the other person is referring to that meeting, or may assume different dates or locations. Coordination depends on the shared interpretation, not just overlapping isolated facts.

Related terms

Grounding
Clarification
dialogue state
reference resolution
conversational repair
shared context
acknowledgment

Learn next

Study conversational repair and definite reference. Rewrite an ambiguous booking exchange to show which details are established at each turn, then test whether a later correction propagates into the final action and summary.

Definite reference

Sources

Herbert H. Clark, research overview: https://web.stanford.edu/~clark/research.html. Common ground and grounding as collaborative processes.

Clark and Brennan, Grounding in Communication: https://web.stanford.edu/~clark/1990s/Clark,%20H.H.%20_%20Brennan,%20S.E.%20_Grounding%20in%20communication_%201991.pdf. Foundational account.

Jurafsky and Martin, Conversation and Its Structure: https://web.stanford.edu/~jurafsky/slp3/26.pdf. NLP dialogue context. The booking examples are fictional.

Previous: Common data model Next: Common noun