Grounding is a property of an answer, not a technique for producing one. An output is grounded when every factual claim in it can be traced back to a specific source you trust, and ungrounded when the model produced it from whatever its weights encode. The distinction matters because fluency is uncorrelated with truth: a model will state a fabricated case citation in exactly the same confident register it uses for a real one.
This gets conflated with RAG constantly, and the conflation causes real damage. RAG is the machinery: retrieve documents, put them in the prompt, generate. Grounding is the outcome you were hoping that machinery would deliver. They come apart all the time. Your retriever pulls three irrelevant passages and the model answers from its training data anyway, producing a confident, sourced-looking, ungrounded response. Or it retrieves the right document and then subtly overstates what the document actually said. Shipping RAG and assuming you have grounding is one of the most common mistakes in production AI.
Verifiability is the load-bearing word. An answer that happens to be true is not grounded if nobody can check where it came from. This is why grounded systems attach citations at the claim level rather than the response level: "the deductible is $1,500 [Policy §4.2]" is checkable, while a paragraph with three sources listed underneath is a gesture at checkability. The strict version, useful in regulated settings, is that the system should refuse to assert anything it cannot cite, and say it doesn't know instead.
Measuring it requires groundedness evals, which work by decomposing an output into individual claims and asking, for each one, whether the retrieved context actually entails it. You end up with two failure counts that need separate fixes: claims with no support in the context (the model invented or recalled them) and claims that contradict the context (the model overrode its sources). The first usually means your retrieval missed; the second usually means your prompt didn't make source precedence explicit enough.
There's a real tension to manage. Hard grounding constraints make a system more trustworthy and less capable. If the model may only assert what's in the retrieved documents, it can no longer connect two facts, apply general knowledge, or reason past what a single passage states, and users experience that as an assistant that keeps saying it doesn't know. Where you set that dial is a product decision that depends on the cost of being wrong. A medical dosing assistant should refuse constantly. A brainstorming tool that refuses constantly is useless.
The pattern that works in practice is layered rather than binary: retrieve well, instruct the model explicitly to prefer sources over its own knowledge and to flag when it is going beyond them, cite at the claim level, and run automated groundedness checks on a sample of production traffic so you find out when retrieval quality drifts rather than hearing it from a customer.
Why It Matters
Grounding is the difference between an AI system that can be deployed into a regulated or high-stakes workflow and one that stays a productivity toy. Legal, medical, financial, and compliance use cases don't just need correct answers, they need answers a human can audit after the fact. It's also the honest answer to "how do we stop hallucination": you don't eliminate it at the model level, you constrain the system so that unsupported claims are visible and checkable.
Example
An insurance claims assistant is asked whether a specific water-damage claim is covered. Ungrounded: "Water damage from burst pipes is typically covered under standard homeowners policies, though sudden and accidental damage is treated differently from gradual leaks." True-ish, generic, and unusable for a decision. Grounded: "Covered. Policy HO-3 §7.1 covers sudden discharge from plumbing systems [source: policy_HO3.pdf, p.14]. The March 3 inspection report classifies this as sudden discharge [source: claim_88231_inspection.pdf]. Note: §7.3 excludes damage from gradual seepage, which the inspector explicitly ruled out." An adjuster can verify every clause in under a minute.