---
title: Grounding
description: 'Connecting an AI output to identifiable evidence, data, observations, or constraints that support its claims.'
canonical_url: 'https://darkfactory.dev/glossary/grounding'
markdown_url: 'https://darkfactory.dev/glossary/grounding.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Grounding


## Definition

Grounding connects a model's output to identifiable evidence, data, observations, or constraints that support its claims. In a grounded answer, a reviewer can trace each material claim to an allowed source and determine whether that source supports the claim.

The evidence may be a supplied document, a database result, tool output, sensor observation, or another authoritative record. Grounding can happen through retrieval, but retrieval alone is only a candidate-selection step. The system still has to preserve provenance, keep untrusted text from becoming instructions, and check that the answer follows what the evidence says.

## Two meanings with a shared history

Stevan Harnad's 1990 paper *The Symbol Grounding Problem* gave cognitive science its classic formulation: how symbols acquire meaning through connection to something other than more ungrounded symbols. Modern LLM engineering uses "grounding" more narrowly for connecting an answer to external evidence or state. The current usage borrows the same connection-to-the-world idea but does not solve Harnad's broader problem.

Google DeepMind's 2024 FACTS Grounding benchmark offers one operational definition for document-based tasks. It asks whether a long-form response answers the request and remains fully attributable to the supplied document. That is stricter than attaching a citation somewhere in the response.

## Operational significance

Grounding is a claim-level property. One supported paragraph does not ground unrelated statements elsewhere in the answer. A source can also be current and reputable without entailing the model's wording. Systems need evidence coverage, entailment checks, and a clear policy for unsupported claims.

## Distinguish it from nearby terms

- Retrieval fetches candidate material. Grounding requires the output to stay supported by the material that should govern the answer.
- Citation provides a pointer. A citation can be real but irrelevant, or relevant but too weak to support the claim.
- Factuality asks whether a claim is true. A response can accurately repeat a false source and be grounded to that source while remaining factually wrong about the world.
- Verification is the process used to test the output. Grounding is one property that verification may examine.

## Check your understanding

A response cites a correct policy document, but the cited section never states the eligibility rule claimed in the answer. Decide whether the response is grounded and identify the missing check.

## Also called

grounded generation

## Related factory areas

- [Verification, evaluation & quality truth](https://darkfactory.dev/factory/verification)

## Evidence and further reading

- [The Symbol Grounding Problem](https://www.cs.ox.ac.uk/activities/ieg/e-library/sources/harnad90_sgproblem.pdf)
- [FACTS Grounding](https://deepmind.google/blog/facts-grounding-a-new-benchmark-for-evaluating-the-factuality-of-large-language-models/)