Definition
Grounding is the practice of anchoring an AI model’s output to verifiable external evidence — retrieved documents, tool or database results, or perceived inputs like images — rather than relying on what the model happens to have memorised during training. A grounded response is one you can trace back to a source and check. Because ungrounded language models generate fluent text regardless of truth, grounding is the primary defence against hallucination: confident but fabricated answers.
Types of grounding
Grounding takes several forms depending on what the output is anchored to:
- Factual grounding — answers are tied to a trusted knowledge source, so claims can be verified against documents rather than the model’s parameters.
- Retrieval grounding — relevant passages are fetched at query time and placed in the prompt; the model answers from those passages.
- Visual grounding — language is linked to specific regions of an image, so a multimodal model can say where “the red car” is in a photo.
Each connects a different kind of model output to something checkable in the world.
How grounding is implemented
The dominant technique is retrieval-augmented generation (RAG): a query is embedded, the most relevant documents are pulled from a knowledge base or vector store, and those documents are inserted into the model’s context with an instruction to answer only from the provided sources and cite them. Tool use and function calling extend this — letting the model fetch live data, run a calculation, or query a database before answering. Adding inline citations makes the grounding visible and lets users audit each claim.
Why it matters
Grounding is what makes AI trustworthy for real work. In legal, medical, financial, and customer-support settings, an unverifiable answer is worse than no answer. Grounded systems can keep knowledge fresh (the source updates, the answer updates) and attributable (every statement points to evidence) — properties that fine-tuning facts into weights cannot easily provide.
Limits and good practice
Grounding reduces but does not eliminate error. A model can still misread a retrieved passage, cite the wrong document, or blend grounded and ungrounded claims. Good practice is to retrieve high-quality, relevant sources, instruct the model to say “I don’t know” when the evidence is missing, surface citations so humans can verify, and evaluate the system on citation accuracy rather than fluency alone.