grounding
A technique that ties a language model’s answer to sources it is given, such as documents or search results, so each claim can be traced to one.
- A search chatbot’s answer with a source link after each sentence
- A contract summary citing the exact clause behind each point
- A flag on a sentence the supplied documents do not support
- Earliest AI use
- 1990
- Symbol grounding: how a system’s symbols can have meaning of their own.
- Current meaning
- 2022
- Google’s LaMDA: replies backed by known sources, scored as groundedness.
- Opposite
- Ungrounded output
- What Microsoft’s detector flags: claims the source material does not back.
In plain words
In AI, grounding is tying a model’s answer to sources it was given, such as documents or search results, so each claim can be checked.
A large language model writes from patterns learned in training, and nothing in that process checks a claim against a source. Grounding supplies material to answer from at the moment of the question, such as documents, search results or a company’s own records. The model is told to base its answer on that material and, often, to cite it.
The benefit is that an answer can be checked. In 2022 Google scored its LaMDA conversation model on groundedness: how often its claims about the world could be backed by authoritative sources. Google noted that grounding did not guarantee accuracy. It let people judge a reply by the reliability of its source.1
The word names the aim, and the methods vary. Retrieval-augmented generation finds relevant passages with a search. A user can also attach a contract, or a service can pass in live web results. Microsoft’s Bing chatbot did this in 2023 to answer questions about recent events with fewer mistakes, and Microsoft called the method grounding.2
Grounding reduces hallucination but does not remove it. A model can misread its sources or state things they do not say.3 It can also repeat a source faithfully when the source is wrong.1 That is the difference from fact-checking, which tests a claim against the world.
In AI research the word is older. From 1990 it described how a system’s symbols could mean something beyond other symbols, a question known as the symbol grounding problem. Some language researchers keep that sense. A 2020 paper argued that shared experience of the world is what makes language meaningful.
How it works
A grounded system works in three stages.
First, it gathers sources and places them in the model’s context window, the text a model can take in at once. The sources may come from a search, a database or a file the user attached. When Google added search grounding to its Gemini API in October 2024, it charged developers $35 per 1,000 grounded queries.4
Second, it instructs the model to answer from those sources and to cite them. Some services return the evidence with the answer. Anthropic’s Citations feature points to the exact sentences and passages that Claude used to write its reply.5
Third, it can check the result. A groundedness check compares the answer with the sources and flags statements they do not support. Microsoft offers one as a service.3 Google DeepMind’s FACTS Grounding benchmark scores models on 1,719 examples with documents of up to 32,000 tokens. Three models from different companies act as judges.6
Origin and first known use
From ground, the solid base beneath something; to ground a claim is to give it a firm basis. The word entered AI in 1990, when Stevan Harnad described the symbol grounding problem. It asks how the symbols a system manipulates can mean something to the system itself, not only to its human readers. In 2017 a knowledge-grounded conversation model drew its replies from outside facts. In 2022 Google’s LaMDA paper measured groundedness as support from known sources. Microsoft, Google and Anthropic built products on that sense from 2023.
- 1990 Stevan Harnad names the symbol grounding problem: how a formal system’s symbols can mean anything to the system itself. He likens it to learning Chinese from a Chinese-only dictionary.Harnad, “The Symbol Grounding Problem”, 1990 arXiv cs/9906002 →: source for 1990
- 2017 Ghazvininejad and colleagues present a knowledge-grounded conversation model that conditions each reply on outside facts as well as the conversation so far.Ghazvininejad, Brockett, Chang, Dolan, Gao, Yih and Galley, “A Knowledge-Grounded Neural Conversation Model” arXiv 1702.01932 →: source for 2017
- 2020 Bisk and 11 co-authors use the older sense in “Experience Grounds Language”, arguing that shared experience of the world makes utterances meaningful.Bisk et al., “Experience Grounds Language”, EMNLP 2020 aclanthology.org →: source for 2020
- 2022 Google’s LaMDA paper names factual grounding as a key challenge and scores it by groundedness: replies backed by known sources, not merely plausible ones.Thoppilan et al., “LaMDA: Language Models for Dialog Applications” arXiv 2201.08239 →: source for 2022
- 2023 Microsoft’s new Bing sends web search results to its chat model, calls the method grounding, and adds citations so users can check sources.Jordi Ribas, “Building the New Bing”, Microsoft Bing blog, 21 February 2023 blogs.bing.com →: source for 2023
- 2024 Microsoft announces groundedness detection in Azure AI, a feature designed to identify text-based hallucinations by checking output against supplied sources.Microsoft Azure blog, “Announcing new tools in Azure AI to help you build more secure and trustworthy generative AI applications”, March 2024 azure.microsoft.com →: source for 2024
- 2024 Google DeepMind launches FACTS Grounding, a benchmark that counts a long answer as accurate only if a supplied document fully supports it.FACTS team, “FACTS Grounding: A new benchmark for evaluating the factuality of large language models”, Google DeepMind blog, 17 December 2024 deepmind.google →: source for 2024
- 2025 Anthropic launches Citations, an API feature that lets Claude point to the exact sentences and passages in supplied documents behind its answers.Anthropic, “Introducing Citations on the Anthropic API”, 23 January 2025 claude.com →: source for 2025
In the wild
-
“…our approach enables the model to generate responses grounded in known sources, rather than responses that merely sound plausible.”
Thoppilan et al., arXiv preprint arXiv 2201.08239 →: Thoppilan et al., arXiv preprint, 20 Jan 2022Archived →: copy of Thoppilan et al., arXiv preprint, 20 Jan 2022 -
“Thanks to the Bing grounding technique, Prometheus is also able to integrate citations into sentences in the Chat answer…”
Jordi Ribas, Microsoft Bing blog blogs.bing.com →: Jordi Ribas, Microsoft Bing blog, 21 Feb 2023Archived →: copy of Jordi Ribas, Microsoft Bing blog, 21 Feb 2023 -
“…the model returns grounding sources (in-line supporting links) and Search Suggestions that point users to the search results corresponding to the grounded response.”
Shrestha Basu Mallick and Logan Kilpatrick, Google for Developers blog developers.googleblog.com →: Shrestha Basu Mallick and Logan Kilpatrick, Google for Developers blog, 31 Oct 2024Archived →: copy of Shrestha Basu Mallick and Logan Kilpatrick, Google for Developers blog, 31 Oct 2024 -
“…responses are judged as factually accurate if they are fully grounded in information contained in the provided document, with no hallucinations.”
FACTS team, Google DeepMind blog deepmind.google →: FACTS team, Google DeepMind blog, 17 Dec 2024Archived →: copy of FACTS team, Google DeepMind blog, 17 Dec 2024 -
“…Citations, a new API feature that lets Claude ground its answers in source documents.”
Anthropic, product announcement claude.com →: Anthropic, product announcement, 23 Jan 2025Archived →: copy of Anthropic, product announcement, 23 Jan 2025 -
“A grounded response adheres closely to the given information, avoiding speculation or fabrication.”
Microsoft, Azure AI Content Safety documentation learn.microsoft.com →: Microsoft, Azure AI Content Safety documentation, 20 Feb 2026Archived →: copy of Microsoft, Azure AI Content Safety documentation, 20 Feb 2026
Not to be confused with
-
Retrieval-augmented generation
One method of grounding: a search step finds passages and adds them to the prompt. Grounding is the aim of tying each claim to a source the model was given, which attached files or live search results can also serve.
-
Fact-checking
Fact-checking tests a claim against the world, using any evidence available. Grounding tests it only against the sources the model was given, so a grounded answer can repeat an error that sits in its source.
-
Symbol grounding
The older AI sense: how the symbols a system manipulates can mean something beyond other symbols. It is a question about meaning, not about whether an answer is backed by a document.
-
Hallucination
The failure grounding aims to reduce: fluent output that is false or unsupported by any source. Grounding lowers the risk without removing it, since a model can still misread its sources or go beyond them.
Questions
What does grounding mean in AI?
Grounding means giving a language model sources to answer from, such as documents or search results, and tying each claim in its answer to them. The sources go into the prompt at the moment of the question, and the answer often cites them. Microsoft’s Bing chatbot used web search results this way in 2023. Google, Microsoft and Anthropic added grounding features for developers in 2024 and 2025.
What is the difference between grounding and RAG?
Grounding is the aim of tying each claim in an answer to a source the model was given. Retrieval-augmented generation, or RAG, is one method for reaching it: a search step finds relevant passages and adds them to the prompt. Other methods include attaching a document to the request or passing in live web search results, as Microsoft’s Bing chatbot did in 2023.
Does grounding stop AI hallucinations?
No. It reduces them, but a grounded model can still misread its sources or state things they do not say. Microsoft built its groundedness detection to catch such errors, and Google DeepMind built a benchmark to measure them. When Google introduced groundedness for LaMDA, it noted that grounding does not guarantee accuracy. It lets people judge an answer by the reliability of its source.
How is groundedness measured?
Groundedness is scored by checking an answer against the sources the model was given. Google’s LaMDA work measured how often replies making claims about the world could be backed by authoritative sources. Google DeepMind’s FACTS Grounding benchmark gives models documents of up to 32,000 tokens and has three other models judge whether each answer is fully supported. Microsoft offers a similar check as a service.
What is the symbol grounding problem?
It is the older AI meaning of grounding. In 1990 Stevan Harnad asked how the symbols a computer manipulates could mean anything to the computer itself, not only to its human readers. He likened it to learning Chinese from a Chinese-only dictionary. In 2020 a group of language researchers argued that shared experience of the world is what makes language meaningful.
Watch
Cite this entry
“Grounding.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Grounding. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Grounding.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/