Generative AI is becoming more and more involved in newsrooms, and the implications of it on framing, sourcing, and trust are under-researched. The paper is a combination of content analysis (politics, health, technology), randomized experiment (n=800), and semi-structured interviews with reporters. The findings indicated that AI-mediated articles were more understandable and coherent but lacked diversity in the sources and were more based on official voices. AI-only stories were otherwise penalized in terms of trust, and AI+editor stories minimized this difference. Interviews highlighted the positive aspects of AI as a speedy and innovative tool, but also mentioned prejudice, illusions, and disclosure expenses. The paper recommends the use of AI alongside editorial values, disclosure, and source variability as a way of ensuring trust. The policy suggestions are to enhance the quality of provenance and to promote diversity of documented sources in AI-assisted journalism.

Generative AI is potentially efficient in the newsrooms, but raises concerns about the issue of credibility and trust. We evaluate its effect and the results of 600 articles each with a stratified content analysis of each production mode (human/AI-assisted/AI-generated) and with disclosure (none/minimal/rich) (1) to determine its effect on accuracy, sourcing, and correction latency; (2) a preregistered 3 × 3 experiment manipulating production mode and disclosure (none/minimal/rich) to determine its effect on perceived article credibility and brand trust. Higher error and hallucination rates and fewer named sources, and slower corrections of AI-generated items are demonstrated by content analysis. Minimal AI labels diminish credibility and trust experimentally, but rich, process-level disclosure, naming, editorial verification, and sources mitigate penalties of work assisted by AI. We give policy and legitimacy implications to the newsroom.