Judge
Grounding & Hallucination Indicators
Can I trust what the system generated?
Distinguishes generated claims that rest on retrievable evidence from those the model may have invented.
Overview
Why it matters
Fluent answers look equally true, whether they are or not.
An AI assistant answers with clean formatting, a confident tone and a plausible citation. Some of it may be accurate and some invented, and the interface shows no difference. On its own, a language model generates text that sounds right, not text it has checked.
This is not the same as Confidence & Uncertainty, which covers error ranges in predictions. Here, a generated sentence may simply not be true.
What does it look like?
Resident checks a generated summary
Resident checks a generated summary
- Resident asks the city’s benefits assistant for a summary of housing support and gets a summary card labeled “AI-generated content, [timestamp]”.
- A badge on the card reads: “3 of 4 statements found in the city’s published pages (checked automatically).”
- Resident opens the badge. Supported lines link to “Housing support rules, city website, updated March 2026”.
- The line “Up to €450 a month” is marked: “Not found in the city’s pages. Ask a caseworker.”
- Resident books a caseworker appointment and asks about the amount.
Citation check
Citation check
- Resident asks the city’s benefits assistant which rule requires the lease for housing support.
- System produces a response with inline citation markers to the city’s pages.
- One citation is marked with a warning icon.
- Tooltip: “This page was not among the retrieved pages. Check it independently.”
- Resident checks the rule with a caseworker before relying on that citation.
“I don’t know” state
“I don’t know” state
- User asks a highly specific question outside the system’s knowledge scope.
- System shows: “I couldn’t find a reliable source to answer this. Here’s some general information (not checked against sources): [general information]. For this specific question, please consult [recommended resource].”
- User is not misled by a fluent but incorrect response.
When does it fit?
Use it when
- A language model writes free-text answers
- Factual errors have real consequences, as in medical, legal, financial or public-service use
- People can’t easily check the claims themselves
Skip it when
- The content is creative, where factuality isn’t expected
- There’s no real check behind it: a fake signal creates false confidence
Design
How do I apply it?
-
Compact first, detail on demand.
Show one grounding state by default. Let people open which claims are and aren’t supported.
-
Tie every signal to a real source.
A confidence score on a made-up answer is meaningless. Grounding means a retrieved document or a checked record.
-
Watch the risky details.
Numbers, names, dates and citations are common failure points and easy to check. Mark them claim by claim.
-
Make “unverified” a normal state.
Not every answer can be grounded. Say so plainly, without treating it as an error.
-
Never show a badge you can’t back up.
A “supported” or “verified” label without a real check behind it can be worse than no label, because it looks trustworthy.
For your technical team
Where retrieval and generation are separate steps, keep two signals apart: retrieval match (how well the sources matched the query, shown as match levels) and generation faithfulness (how closely the answer reflects those sources). Research distinguishes intrinsic hallucination, which contradicts a source, from extrinsic hallucination, which can’t be checked against the source; use different markers for each. Automated support checks (NLI models or LLM judges, as in FActScore) are estimates and make errors, so call their result “supported” or “found in sources” and keep “verified” for checks a person made.
How can I show it?
Grounding status badge
Supported, partly supported or not found, per answer.
A compact label attached to a response or section indicating its grounding state: - Supported: claims supported by retrieved sources (checked automatically) - Partly supported: some claims supported, others not found - Not found in sources: response generated without a supporting source - Potentially outdated: sources retrieved are older than a threshold. Keep “verified” for checks a person made.
Recency indicator
When the newest supporting source was last updated.
For time-sensitive domains, a signal showing when the most recent supporting source was last updated, helping users assess whether the information is still current.
I don’t know design state
A designed way for the system to say it doesn’t know.
An explicit interface state for when the system has low grounding confidence and should say so, not produce a fluent but potentially wrong response. This is a product design decision, not just a model decision.
Claim-level verification markers
Marks individual sentences, not the whole answer.
Inline highlights or markers on specific sentences or claims, not the entire response. Use different markers for “contradicts a source” (intrinsic hallucination) and “not found in any source” (extrinsic hallucination), because they call for different next steps. Set how prominent each is by what’s at stake in the claim, not by type: an invented amount or dosage needs a strong marker even if no source contradicts it.
Source mismatch warning
Flags claims that contradict or don’t appear in any source.
A visible warning when the generated response makes a claim that contradicts a retrieved source, or when the response contains specifics (dates, numbers, names) that don’t appear in any retrieved document.
Confidence-grounding split
Separates how well sources matched from how faithful the answer is.
In systems that have both a retrieval step and a generation step, separating two signals: - Retrieval match: how well the sources matched the query - Generation faithfulness: how closely the output reflects what was retrieved. These are different things, and conflating them can mislead users.
What does it cost?
Transparency vs. paralysis
Warnings on every answer stop being read. When everything is flagged, nothing stands out. Save visible signals for unsupported or high-stakes content.
Grounding vs. false assurance
“Grounded” means consistent with a source, not that the source is right or current. An automated check can also be wrong. Say what the signal means and how it was checked.
Flagging vs. refusing
Some systems answer with a warning, others refuse. Answering can mislead, and refusing can leave people without help. Set the threshold on purpose, and offer a person when the system holds back.
Connections
Which patterns work with it?
Confidence & Uncertainty
How certain is the system?
Uncertainty addresses probabilistic prediction confidence. Grounding addresses factual faithfulness in generative systems. Related but distinct failure modes that need different design treatments.
Retrieved Source Attribution
What sources did the system draw from?
Attribution shows what was retrieved. Grounding signals how faithfully the output reflects it. Attribution without grounding tells users the source but not whether it was accurately represented.
Model Scope & Limits
What can this system reliably do?
Scope defines where the system applies. Grounding checks each answer, even within scope.
Inspection Dialogue
Can I ask the system follow-up questions?
“How do you know this?” is a natural grounding question that this pattern lets users ask.
Sources (3)
Sources (3)
Ji et al. (2022) — Survey of Hallucination in Natural Language Generation
comprehensive taxonomy of hallucination types in NLP systems (ACM Computing Surveys, 2023); distinguishes intrinsic hallucination (contradicts the source) from extrinsic hallucination (can’t be checked against the source). A basis for designing grounding signals at the claim level
Min et al. (2023) — FActScore: Fine-Grained Atomic Evaluation of Factual Precision in Long Form Text Generation
decomposes generated text into atomic claims and evaluates each claim individually against a knowledge source, rather than scoring the response as a whole. Establishes the technical basis for claim-level grounding indicators rather than response-level confidence scores. Published at EMNLP 2023
Es et al. (2023) — RAGAS: Automated Evaluation of Retrieval Augmented Generation
proposes automated metrics that separate faithfulness from retrieval relevance. Useful for checking a pipeline before showing users any grounding signal
Your AI tool answers confidently, and nobody can tell when it’s making things up.
That matters most in legal, medical, compliance or research work. That’s where this pattern starts. Let’s look at your interface together.
Work with me