Glanceable Verification Indicators for AI-Assisted Storytelling

AI-assisted knowledge workers face challenges, including over-reliance on potentially incorrect outputs and LLM hallucinations. Manual verification requires time and domain expertise, and is particularly problematic under deadline pressure. We investigated whether visual confidence cues could serve as glanceable verification indicators in AI-assisted climate storytelling. In a within-subjects study (n = 26), creative professionals edited scripts with AI assistance. We compared baseline and verifiable conditions, both providing AI-generated inline quotes with numbered citation badges. The verifiable condition added color-coded confidence indicators, which matched quotes to sources (green for strong, yellow for partial, and red for weak match). Initial results showed that despite both conditions rating similarly for perceived verification ease, the verifiable condition led to higher trust ratings, citation hovers, and source views. We contribute glanceable verification indicator designs with empirical evidence of increased verification engagement and perceived trust in AI-assisted storytelling, with design implications for verifiable AI systems.

• https://dl.acm.org/doi/full/10.1145/3772...