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.
- Evidence: research-based (Ji et al. 2022, Min et al. 2023)
- Tested: concept only, mock data
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 is accurate, some is invented, and the interface shows no difference. 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?
High-stakes fact check
High-stakes fact check
- User asks a medical question.
- System responds with a grounding badge: "Partially grounded: 2 of 4 claims verified."
- User clicks the badge.
- Panel expands showing which sentences are highlighted as verified (green) and which are unverified (amber).
- User sees a warning: "The specific dosage mentioned could not be verified in retrieved sources."
- User consults a verified reference or medical professional.
Citation check
Citation check
- User asks for a summary of a research area.
- System produces a response with inline citation markers.
- One citation is marked with a warning icon.
- Tooltip: "This citation could not be verified in the retrieved documents. Check independently."
- User removes the suspect citation from their own work.
"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 what I do know: [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 or financial 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 verification behind it: a fake signal creates false confidence
Design
How do I apply it?
-
Tie every signal to a real source.
A confidence score on a made-up answer is meaningless. Grounding means a retrieved document or a verified record.
-
Make "unverified" a normal state.
Not every answer can be grounded. Say so plainly, without treating it as an error.
-
Watch the risky details.
Numbers, names, dates and citations are the most often invented. Mark them claim by claim.
-
Never show a badge you can't back up.
A "verified" label without real verification is worse than no label.
-
Compact first, detail on demand.
Show one grounding state by default. Let people open which claims are and aren't supported.
For your technical team
Where retrieval and generation are separate steps, keep two signals apart: retrieval confidence (how well the sources matched the query) 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 any source. They need different treatments.
How can I show it?
Grounding status badge
Grounded, partly grounded or unverified, per answer.
A compact label attached to a response or section indicating its grounding state: - Grounded: claims verified against retrieved sources - Partially grounded: some claims supported, others not - Unverified: response generated without supporting source - Potentially outdated: sources retrieved are older than a threshold
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.
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.
Claim-level verification markers
Marks individual sentences, not the whole answer.
Inline highlights or markers on specific sentences or claims, not the entire response. Hallucination research distinguishes two types that warrant different visual treatments: - Intrinsic hallucination: the response contradicts a retrieved source. Higher severity; warrants a strong warning marker. - Extrinsic hallucination: the response contains claims that cannot be verified against any retrieved source, neither supported nor contradicted. Lower certainty; warrants an "unverified" marker rather than a contradiction warning. This distinction matters for design: a contradiction signal and an "unverifiable" signal look similar if treated identically, but carry very different implications for what the user should do next.
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 confidence: how well the sources matched the query - Generation faithfulness: how closely the output reflects what was retrieved These are different things and conflating them misleads users.
Further left for everyday answers, further right for high-stakes domains like legal or medical.
What does it cost?
Transparency vs. paralysis
Warnings on every answer stop being read. Save visible signals for uncertain or high-stakes content.
Grounding vs. false assurance
"Grounded" means consistent with a source, not that the source is right or current. Say what it means.
Flagging vs. refusing
Some systems answer with a warning, others refuse. Both are valid. Set the threshold on purpose.
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 requiring 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. These two patterns work together: 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 signals when a response is outside the system's reliable knowledge 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 enables users to ask.
Sources (4)
Sources (4)
Ji et al. (2022) — Survey of Hallucination in Natural Language Generation
comprehensive taxonomy of hallucination types in NLP systems; distinguishes intrinsic hallucination (contradicts source) from extrinsic hallucination (unverifiable against source). Foundational 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
introduces faithfulness as a measurable dimension separate from retrieval relevance. The faithfulness metric is the technical equivalent of what this pattern surfaces to users
Kroll et al. (2017) — Accountable Algorithms
while predating LLMs, establishes the accountability principle that systems making consequential decisions must be auditable and their reasoning examinable. Grounds the normative case for surfacing hallucination risk in high-stakes contexts
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