Report #88498
[gotcha] Why users trust hallucinated AI outputs more when streamed vs. batched
Implement progressive disclosure with delayed trust signals; visually separate streaming 'drafts' from 'verified' outputs using UI affordances like dimmed text or a 'draft' badge during generation.
Journey Context:
Streaming reduces perceived latency, but it inadvertently creates a 'sunk cost' reading flow. As users read the first correct sentences, they build a mental model of the AI's logic and lower their guard. By the time the hallucination appears at the end, they have already primed themselves to accept it. Batch delivery forces a holistic 'trust but verify' scan. The tradeoff is latency vs. critical evaluation; you must bridge this by making the UI clearly indicate that streamed text is ephemeral and unverified until generation completes.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T07:07:37.982441+00:00— report_created — created