Can readers tell truth from fabrication without evidence signals?

The paper reports that readers given no provenance signal could not separate truth from fabrication. In the control condition, high-fluency hallucinations averaged M = 6.28 and ground truth M = 5.78 (p = .43). The excerpt does not state the rating scale behind those means. When an idealized Provenance Density interface was shown, the authors report a discernment gap of +4.15 points (d = 1.82, p < .001) between truth and fabrication. They call the control result a "Fluency Trap": readers trust fluent text as if fluency still cost something to produce. Participant comments fit that reading. One said the claims "sound believable and well written," and another judged accuracy by "how the text was worded and flowed."

The authors' mechanism runs through costly signaling. Polish once worked as a handicap, since faking it was expensive, so it separated competent writers from others. Generative AI collapses that separating equilibrium into a pooling one, where expert testimony and fabrication share the same form. Provenance Density is their response: a score built from atomic claims, each checked by retrieval and weighted by source reputation and relevance, displayed as the density of verified claims. Binary "Made with AI" labels answer "Who wrote this?" rather than "What supports this?" The excerpt says binary labels "acted as a blunt warning," and a participant called the banner "a warning vs being informative." The authors read the label as a risk cue, an accuracy-independent discount they call the Transparency Penalty. The excerpt gives no statistics for the label condition.

This extends Does polished AI output trick audiences into trusting it? from charts and decks to prose, and tests the reader side. The presentation authority that note describes is what the no-signal participants appear to have leaned on. It also gives the fluency illusion in How do AI tools trick users into overestimating their own skills? a measured counterpart. That note concerns a producer's inflated self-assessment, while this paper shows fluency standing in for truth in a reader's judgment. The paper also sits beside Do users worldwide trust confident AI outputs even when wrong?. That note locates the problem in confidence cues. The paper tests a verified-claim cue and does not test confidence markers, so it cannot say whether evidence density would correct that overreliance.

The result is an upper bound by design. The Oracle protocol showed participants idealized values, with grounded summaries paired with high-density indicators and fabricated ones with null indicators, so the user study does not test the live pipeline. The technical audit (N = 200 on a composite of TruthfulQA and FreshQA) found that "retrieval density alone is insufficient," and the authors list retrieval reinforcing popular misconceptions as a live risk. The Latin-square design did not fully cross interface, veracity and topic; the Provenance Density–hallucinated cell was measured on one topic, Matcha. The score is also conservative by design: established facts averaged 0.79 and emerging dynamic topics 0.64. The evidence supports that a correct evidence-density display can change how people discriminate. It does not show that a deployed pipeline produces correct densities, or that the effect holds across topics. Treat provenance density as a tested direction for transparency, with high-density false positives as the open risk.