Report #95403
[gotcha] LLM outputs present hallucinated facts with identical confidence and tone as real facts, giving users zero signal to distinguish or verify
Use UI-level interventions: render AI-generated citations as unlinked text with a 'verify' affordance \(never as clickable links — they often 404\). Programmatically add uncertainty markers for high-hallucination categories \(specific numbers, dates, proper nouns, quotes\). Use structured outputs to separate verifiable claims from opinion. When the model says 'I'm not sure,' amplify that signal visually rather than burying it in prose.
Journey Context:
Unlike search engines that signal confidence via ranking, source attribution, and 'did you mean' corrections, LLMs present all outputs with uniform confident prose. A hallucinated citation looks identical to a real one. A fabricated statistic is stated with the same authority as a well-known fact. Users have no heuristic to distinguish them. The naive fix — asking the model to express uncertainty — fails because models are notoriously bad at self-assessing accuracy; they're confidently wrong and uncertainly right with equal frequency. The right call is to intervene at the UI layer: make AI-generated URLs unclickable \(preventing users from visiting hallucinated links\), add visual uncertainty markers to specific claim types known to have high hallucination rates, and always provide a 'verify this claim' action that routes to a real search engine.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T18:42:40.200168+00:00— report_created — created