Provenance Density Interface Yields +4.15 Point Discernment Gap in 81-Participant Study

The paper defines the Fluency Trap as user over-trust in fluent AI output combined with discounting of accurate AI-generated content under binary disclosure. It replaces authorship labels with a density visualization of verified claims backed by retrieval and consistency checks. A technical audit of 200 samples found retrieval density alone insufficient; the Consistency Veto supplied most discriminative power on dynamic queries.

The study isolates evidence visualization from authorship signals. Participants shown only density maps separated truth from fabrication at scale; controls without the map could not. This directly tests the transparency penalty described in the abstract and extends prior work on citation interfaces in retrieval-augmented generation systems.

Operationally the result implies platforms must instrument claim-level provenance rather than post-hoc labels. Current binary tags fail to surface the Consistency Veto signal the audit identified. Deployment requires query-time retrieval pipelines that expose veto decisions, not static watermarks.

Next steps include live A/B tests on production summarization systems and measurement of veto false-positive rates under distribution shift. The 2026 arXiv submission supplies the first controlled evidence that density maps close the discernment gap where disclosure alone does not.