Choose source grounding display when the product can show answer-wide source scope, searched sources, retrieved sources, used evidence, unsupported claims, partial support, permission-limited sources, stale sources, and retrieval failures.
Choose source grounding display when the product can show answer-wide source scope, searched sources, retrieved sources, used evidence, unsupported claims, partial support, permission-limited sources, stale sources, and retrieval failures.
Practice the judgment: Imagine this situation: Retrieval fails but the UI hides the failure and returns a confident model-memory answer. Write the person's goal, choose one of these patterns, and show the state that helps the person continue. Use the decision rules and source trail below to explain your choice.
Review the underlying principle →| Dimension | AI answer without sources | Source grounding display | Citation display | Confidence / uncertainty display | Streaming response | Correction feedback |
|---|---|---|---|---|---|---|
| UI or UX | UI + UX - Unsourced AI answer anti-pattern | UI + UX - Whole-answer source coverage and grounding evidence display | UI + UX - Inline claim-to-source evidence display for generated or summarized content | UI + UX - Calibrated reliability and uncertainty display for AI or automated predictions before user action | UI + UX - Incremental generated-output surface that renders response chunks before final completion | UI + UX - AI output correction feedback flow with selected claim, reason, scope, consent, receipt, and review state |
| UI guidance | Give each control a clear label and visible state. | Render source grounding as an answer-wide evidence panel that separates source scope, searched sources, retrieved sources, used sources, supported claims, partially supported claims, unsupported claims, and unresolved source states. | Render citation markers beside the claims they support, and connect each marker to a selected source preview with title, source type, excerpt, date or version, permission state, and open source action. | Render confidence and uncertainty as labelled reliability information with confidence band, reason, input scope, calibration status, review threshold, freshness, and the next safe action. | Render streamed output with a visible generation state, partial-answer label, stop control, final-complete state, and clear distinction between text that is still arriving and content that has passed final citation, safety, tool, or format checks. | Render correction feedback as an AI-output control that anchors feedback to a specific answer, claim, source, recommendation, tool result, or message instead of a detached thumbs signal. |
| UX guidance | Help users know whether they can rely on an AI answer by exposing its evidence state before they act. | Let users inspect coverage gaps, source scope, retrieval status, and unsupported claims before they copy, publish, apply, or rely on the answer. | Keep the claim-to-source mapping durable through streaming, regeneration, copying, mobile layout changes, and permission checks so users can tell what evidence applies to which claim. | Use confidence / uncertainty display when users need to decide whether an AI prediction, classification, recommendation, extraction, risk assessment, or generated answer is reliable enough to apply. | Use a streaming response when showing partial generated output helps users start reading or monitoring work before the model finishes, and when the product can explain that early chunks may still change, be filtered, or lack final sources. | Separate correction capture from regeneration and manual editing: users should know whether their feedback updates this answer, future answers, reviewer queues, model improvement data, or none of those. |
| Good UI | The interface shows a clear action and state. | A policy answer includes a Grounding panel showing 4 sources searched, 3 retrieved, 2 used, 5 supported claims, 1 partially supported claim, and 1 unsupported claim with a Review action. | A policy assistant places numbered citation chips after each sourced claim; selecting a chip opens a source preview with the document title, section, quoted excerpt, updated date, and Open source action. | A claim classifier says Medium confidence, 71 to 78 percent calibrated range, review threshold 80 percent, conflicting account-age signal, and routes the case to manual review. | A policy assistant shows Answer generating, streams paragraphs into a stable answer region, marks citations pending, exposes Stop generation, then changes to Complete when citations and safety checks finish. | An assistant answer lets a user select a suspect claim, choose Wrong answer, add the expected answer and corrected source, opt out of training use, and see Feedback received with review routing. |
| Bad UI | A retrieval failure is hidden and the model returns an answer from prior knowledge. | The answer shows a green Grounded badge even though only one citation supports one paragraph. | A citation appears verified while retrieval is still pending. | Users over-trust an AI answer because a confidence badge looks authoritative. | A streaming JSON answer is displayed as valid data before the closing braces and schema check have completed. | The feedback form captures only sentiment and loses the correction reason or expected answer. |
| Good UX | A user sees that a customer-facing answer is unsourced, runs source search, reviews one unsupported claim, then sends only after the source state changes. | The person reaches the intended result with clear feedback. | A user checks a claim, opens its source preview, compares the quoted excerpt with the answer text, and copies the citation with the source title included. | A reviewer sees low confidence and out-of-distribution input, opens the reason panel, collects the missing invoice, and avoids auto-denying the claim. | A user sees the first-token state quickly, reads early outline bullets while the answer continues, stops generation after enough detail, and sees the result labelled Partial with Continue and Regenerate options. | A user flags a wrong eligibility answer, attaches the correct policy source, applies the correction to the current answer only, and receives a reviewer-routed status. |
| Bad UX | A user sends an unsourced renewal-policy answer to a customer because the fluent answer looked complete. | Grounded status is shown before retrieval finishes. | A user trusts a generated compliance claim because it has a number beside it, but the number points to an unrelated source. | A model update invalidates old confidence scores but the UI keeps showing them as current. | A user copies an early legal recommendation before the final paragraph reverses the conclusion after a tool result arrives. | A user reports that a cited source is wrong, but the feedback is stored as an undifferentiated dislike and future answers keep using the same bad source. |
| Best fit | Users may rely on the answer for policy, compliance, customer support, operations, finance, legal, medical, safety, or public content. | Users need answer-wide evidence coverage before trusting generated content. | Users need to verify generated claims, summaries, recommendations, or extracted facts against source material. | Users must judge whether an AI prediction, classification, recommendation, extraction, risk score, or generated answer is reliable enough to use. | Generated text or structured content can be read or monitored before completion. | Users can identify wrong, unsupported, stale, unsafe, biased, irrelevant, or wrongly sourced AI output after it is generated. |
| Avoid when | The answer already includes valid claim-level citations and source previews; use citation display for improvements. | The only task is opening one claim's source preview, which belongs to citation display. | Links are simply optional onward reading rather than evidence for the answer. | Users only need source coverage or claim evidence, which belongs to source grounding display or citation display. | Intermediate chunks may expose unsafe, private, or misleading content. | Users are editing a generated draft directly rather than reporting what the AI got wrong. |
| Required state | Sourced answer state with citations or grounding visible. | Default grounded state with source scope, searched sources, retrieved sources, used sources, and supported-claim count. | Default answer with cited claims and inline citation markers. | High confidence state with calibration scope, reason, and whether direct apply is allowed. | Queued or receiving state before first output arrives. | Initial AI answer state with feedback entry points attached to the answer, individual claims, citations, recommendations, or tool-derived values. |
| Accessibility burden | Make source search, retry, upload source, request access, ask human, and draft-only controls keyboard reachable. | Expose grounding summary, source scope, status counts, unsupported claims, and source groups as text. | Give citation markers accessible names that include their selected state and source status, such as Citation 2, verified source, or Citation pending. | Expose confidence label, uncertainty reason, threshold, freshness, and gated action as text rather than relying on color, position, or animation alone. | Expose stream milestones such as started, still generating, stopped, failed, citation ready, and complete as status messages. | Expose selected claim, selected source, reason, scope, privacy choice, submission state, receipt, review state, and outcome as text. |
| Common misuse | Adding a generic disclaimer while still letting users copy or send the unsourced answer as if it were verified. | Showing a global Grounded badge when only some claims have evidence. | Displaying a link dump below the answer instead of mapping sources to specific claims. | Showing a fake percent or exact decimal for an uncalibrated model score. | Letting partial output look finished before the final event arrives. | Using a thumbs-down or smiley reaction as the only mechanism for correcting wrong AI output. |