Safety-first generative UX: how a search startup balanced helpfulness and hallucination risk

AI · 6 min read

The startup had built a generative search assistant that indexed customer docs and returned concise answers. Early users praised the speed but reported inaccurate assertions and missing citations. Designers and engineers faced a core product decision: maximize fluency and brevity or prioritize transparency and verifiability.

The team iterated on a hybrid card design that shows a short model summary with a visible confidence meter, followed by an expandable, paginated list of exact source excerpts and document links. They also added an “edit-answer” affordance so users could correct or annotate responses, creating a lightweight provenance log.

In A/B tests across 12,000 queries, the hybrid UX lowered reported hallucinations by 67% and raised a “trust” score in NPS-linked surveys. Time-to-answer increased slightly for power users but decreased for new users grateful for clear sources. The case highlights how product design can mediate model risks without sacrificing the speed advantage of generative systems.