Mattis, N., Kieslich, K., & de Vreese, C. H. (2026). Feeling iffy about generative AI: Investigating audiences’ trustworthiness perceptions of task-specific AI disclosures. Digital Journalism, 1–21. https://doi.org/10.1080/21670811.2026.2703599
Summary
This paper examines how disclosing generative AI use in journalism affects audiences’ trust, moving beyond the crude “AI-generated” vs. “human-generated” labels of prior work to test task-specific disclosures across the news value chain. Using a preregistered conjoint experiment with 683 Dutch respondents, the authors show that every AI disclosure lowers perceived trustworthiness, but that the size of this penalty depends on which task the AI performed and on individual reader characteristics. The core argument is that transparency about AI in newsrooms carries a consistent trust cost, yet this cost is heterogeneous and cannot be captured by generic labels — and that this cost should not be used as an excuse to abandon meaningful disclosure.
Key Contributions
- Empirical evidence on nuanced, task-specific AI disclosures rather than binary AI/human labels.
- Identification of individual-level moderators (political position, knowledge of journalistic AI) shaping disclosure effects.
- A typology of five audience preference profiles with distinct socio-demographic and attitudinal predictors, offering practitioners guidance for tailoring transparency.
- Demonstration that the negative disclosure effect is robust across news topics of varying controversiality.
- Engagement with normative debates on what “meaningful transparency” should look like in journalistic workflows.
Methods
A preregistered conjoint experiment (N = 683, quota sample broadly representative of the Dutch population) presented mock news articles accompanied by an “AI Monitor” table disclosing whether each of seven journalistic tasks (idea generation, image generation, background research, article writing, proofreading, fact-checking, and human-in-the-loop) was performed by a human or AI. A fractional factorial design yielded eight stimulus compositions, and respondents rated 24 profiles on a 7-point trustworthiness scale across three pretested news topics (vaccination, housing, logistics). The authors computed AMCEs and AMCIEs (with human performance as baseline) using the cjoint package, tested political position and AI journalism knowledge as moderators, and ran an exploratory k-means cluster analysis (k=5) on respondents’ regression coefficients, followed by MANOVA/ANOVA and Tukey HSD tests to profile cluster membership.
Findings
- Every AI disclosure significantly reduced trustworthiness, by 0.23–0.62 points on a 7-point scale.
- A clear task hierarchy emerged: idea and image generation carried the smallest penalties, production tasks larger ones, and fact-checking the largest; missing a human-in-the-loop had only a modest effect.
- No significant differences across the three news topics of varying controversiality.
- Politically left respondents reacted more strongly against AI-written articles; right-leaning respondents were less concerned about AI fact-checking.
- High-knowledge respondents reacted more negatively to AI fact-checking across all topics.
- No moderator reversed the overall negative direction — disclosure never increased trust for any subgroup.
- Five preference clusters (for vaccination): Cautious Optimists (27.7%, the only group generally trusting AI), Indifferents (28.7%), Fact-Checkers (15.2%), Human Creatives (14.2%), and Human in the Loops (14.2%).
- Least concern about AI use appeared among respondents with the least factual knowledge of how journalists actually use GenAI.
Connections
This work sits at the intersection of public perceptions of AI and synthetic-media transparency, sharing concerns with other studies of how audiences respond to AI disclosure and labeling in news, such as Hameleers2026-mc and Dierickx2026-tw. Its emphasis on audience trust and heterogeneous perceptions of AI in journalism also relates to broader work on generative AI’s role in news and information ecosystems like Suk2026-ai and Gilardi2026-hw.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify