Groundedness: what separates a supported answer from a plausible one
In short
Groundedness is a relation between a claim and the evidence supplied with it, not a synonym for correctness. An answer is grounded when every claim traces back to a passage placed in front of the model. That gives 4 states rather than 2, and the expensive one is true but unsupported: it passes review, cannot be checked, and fails silently the day the fact moves.
Key takeaways
- Grounded is about the evidence supplied. Correct is about the world. Either can hold without the other.
- The dangerous cell is true and unsupported: it survives review and breaks the day the fact changes.
- Groundedness is judged against what was retrieved, not against what exists somewhere in the corpus.
- A citation on a paragraph does not establish support for a specific number inside that paragraph.
An answer is grounded when each claim traces to a passage supplied with the question. That is the whole definition, and the important part is what it omits: nothing about whether the claim is true, useful or well written. Groundedness relates a sentence to its evidence; correctness relates it to the world.
Keeping the 2 apart is not pedantry: a system can only check one of them. You can verify mechanically that a sentence rests on a retrieved passage; you cannot verify from inside the system that it is true. Every practical assurance for a retrieval assistant is built on the first.
Four states, not two
| Relation to the supplied evidence | The statement is true | The statement is false |
|---|---|---|
| Supported by a retrieved passage | The target state. Checkable by a reader in seconds, and still checkable later | The evidence itself is wrong or superseded, and the answer repeats it faithfully |
| Not supported by anything retrieved | Right from training data or luck. Unverifiable, and silently wrong when the fact changes | The plain fabrication, and the only cell most teams ever test for |
The bottom-right cell gets the attention and is the least interesting: it is the one people already look for. The other 2 are where systems lose trust. An answer faithfully repeating a policy replaced last quarter is grounded and wrong — a corpus problem wearing a generation costume, diagnosed in an assistant citing a superseded policy.
Why true and unsupported is the expensive one
A true answer with no supporting passage passes every human review it will get. The reviewer knows the domain, reads the sentence, agrees, moves on. Nothing reveals that the system read it nowhere, so nobody learns this question is not answerable from the corpus.
The bill arrives later. When the fact changes, the corpus is updated and the answer is not, because the answer never came from the corpus. A build that tolerates ungrounded-but-correct output has made those answers unmaintainable, with no signal saying which ones they are.
What has to be true for a claim to be checkable
- Claims have to be atomic. One sentence carrying a figure, a date and a condition is 3 claims, and support for the sentence is not support for the figure inside it.
- Citations have to land on the span, not the document. Pointing at a 4-page policy is not evidence for a number in its third table — the failure mode in a citation that lands on the wrong paragraph.
- Negative claims need different evidence. There is no clause about assignment cannot rest on any single passage: it is a claim about the whole corpus, and needs a result set rather than a quotation.
- Aggregates inherit every gap beneath them. A total across 5 retrieved passages is only as grounded as the least complete of them, and completeness is not something top-k retrieval promises.
A true answer nobody can trace is a maintenance liability that reads like success.
Where support tends to break first
- Questions needing 2 dependent lookups. The model retrieves the first document and fills the second from memory — the structure defined in what makes a question multi-hop.
- Claims about relationships. Who reports to whom, which entity owns which contract: the fact lives between documents rather than in one, which is the case for relationships beating similarity between passages.
- Vocabulary mismatch. The supporting passage exists and never surfaces, so the model answers from prior knowledge — the split in lexical and semantic retrieval failing differently.
- Questions with a confident public answer. Anything the model plausibly saw in training is where it fills a gap without noticing it did.
What groundedness is not
It is not accuracy, and one number reported for both measures neither. It is not confidence, a property of the model rather than the evidence. And it is not the presence of citations: an answer can carry 3 links and assert something none of them says, which makes citation display a usability feature rather than a guarantee.
It is also not stable across models. The same passages and the same prompt produce differently grounded answers from different providers, and one leaning harder on prior knowledge looks better on correctness and worse on support — something to re-check per provider rather than assume, as running one inference provider or several sets out.
How to measure it — sampling, rubric, who or what judges — belongs to evaluation, not to this definition. The consequence belongs here: where a system acts on its answers rather than displaying them, an ungrounded claim stops being a bad sentence and becomes a bad action, which is why we treat support as a design constraint in AI agents and automation. What counts as adequate support is a product judgement that stays with whoever operates the system, and the shapes that responsibility takes are compared in AI product studio vs software agency vs IT consultancy. Neighbouring terms sit in retrieval and grounding, part of the engineering library.
Frequently asked questions
Short answers to the follow-ups this page tends to raise.
What does grounded mean in AI answers?
It means every claim in the answer traces to a passage supplied with the question. Grounding relates text to its evidence, so it says nothing about whether the claim is true. An answer built entirely from retrieved passages is grounded even if those passages are 3 years out of date.
What is the difference between groundedness and correctness?
Groundedness compares a claim to the evidence supplied; correctness compares it to reality. They are independent, producing 4 cases: true and supported, true and unsupported, false and supported when the source is wrong, and false and unsupported. Only groundedness can be checked inside the system.
Is an answer ungrounded if the supporting document exists but was not retrieved?
Yes. Groundedness is judged against the passages placed in the model's context, not everything in the corpus. That is what makes it useful: an ungrounded answer says the evidence never arrived, which points at retrieval rather than the prompt.
Do citations prove an answer is grounded?
No. A citation shows a document was retrieved, not that it says what the sentence claims, and an answer can carry 3 links while asserting something none of them supports. Citations become evidence only when they resolve to the span a claim rests on, which is a wiring decision rather than a display one.
- groundedness
- attribution
- generation
- definitions
The work behind this page
Builds from our portfolio that this page draws on.
AskVault
An AI internal knowledge-search platform that answers employee questions from your own docs — grounded in citations, with knowledge gaps surfaced and deflection tracked.
Productivity AIPriorPilot
An AI prior-authorization and denial-management platform that auto-assembles and submits auths, predicts denials before submission, and drafts the appeals to recover revenue.
Healthcare AIRead next
- The right passages were retrieved and the answer is still wrongIf the answer-bearing passage was in the context and the answer is still wrong, retrieval tuning cannot help. Four synthesis defects produce this, and each leaves its own mark in the output.diagnostic
- The model answered from memory and ignored the passage you gave itPlant a passage that contradicts common knowledge, then ask a question only that passage answers. Whichever the answer follows tells you whether you have an override or a payload that never arrived.diagnostic
- The citation points at the wrong paragraphA misaligned citation is a pipeline bug with a deterministic cause; an invented one is a grounding failure. One exact-string search against the original file separates them in under 5 minutes.diagnostic
- The assistant quoted a policy that was replaced last quarterThe old version is still indexed, or it is gone and a cache is still serving it, or both versions are live and the old one ranks higher. One probe separates them.diagnostic
- Adding one new source made unrelated answers worseNothing about the old passages changed. A new source ranks moderately well on a great many queries, and a fixed top-k has to give it those slots by taking them from something else.diagnostic
- Chunk overlap: what it protects against and what it duplicatesOverlap is insurance against a chunk boundary landing in the middle of one idea. The premium is paid in duplicated candidates crowding a fixed number of prompt slots.definition
Working on something in this space?
Tell us where you are in a sentence or two. We'll tell you honestly whether we're the right team, and what a sensible first slice of the work looks like.
Start the conversation