Grounding in AI

Grounding means that an AI model bases its answer on sources you give it, such as your own documents, a product database or a web search, instead of only on what it picked up during training. A chatbot on a webshop then answers from your actual delivery terms and can show the passage it used, so an employee or a customer can check the answer.

What does grounding mean for language models?

Google Cloud defines grounding as the ability to connect model output to verifiable sources of information. When a model has access to specific data sources, its output stays tied to that data and it is less likely to invent content. Google names fewer hallucinations and links to the sources as the benefits. The word is older than language models: in 1990 the cognitive scientist Stevan Harnad described the symbol grounding problem, the question of how the symbols in a computer system get a meaning of their own, apart from the meaning people give them.

How does grounding work?

Before the large language model answers, the system adds the relevant sources to the prompt. Common ways to find them:

• A search in your own documents: RAG (retrieval-augmented generation).
• A web search. Google offers grounding with Google Search for its Gemini models.
• Live data from a system, such as stock levels or an order status, fetched through function calling.
• Facts from a knowledge graph.

Two further steps make an answer checkable. With citations, the model points to the passages it used: Anthropic's Citations feature (January 2025) splits documents into sentences and has Claude refer to the exact sentences behind each claim. A check afterwards compares the answer with the sources: Microsoft's groundedness detection flags statements that the source material does not support. According to its documentation (last revised November 2025), it handles English text only.

Does grounding prevent hallucinations?

It reduces them. In a human evaluation by Shuster et al. (Facebook AI Research, 2021), a chatbot model without retrieval hallucinated in 68.2% of its answers on Wikipedia topics it had not seen in training; with retrieval, the share was between 7.9% and 21.5%, depending on the setup.

It does not end them. The makers of legal research tools from LexisNexis and Thomson Reuters, which look up case law before they answer, advertised them as avoiding hallucinations or giving "hallucination-free" citations. In a preregistered study, Magesh et al. (2024) found that these tools still hallucinated in 17 to 33% of cases. The search can return the wrong passages, and the model can still add claims that its sources do not support. Google DeepMind's FACTS Grounding benchmark (January 2025) measures this: models get a document of up to 32,000 tokens and a request, and several AI judges check whether the long answer is fully supported by the document.

Example

Suppose an accounting firm lets clients ask about tax deadlines in a chat window. Without grounding, the model answers from its training data, which may be out of date. With grounding, the system first searches the firm's own current overview of deadlines, gives the matching passage to the model with the question, and shows that passage under the answer. When the search finds nothing relevant, the bot says so and refers the client to an adviser.