Gilardi, F., Lorenzo, S. D., Ezzaini, J., Santa, B., Streiff, B., Zurfluh, E., & Hoes, E. (2026). Quality perceptions and intended engagement in response to AI-generated and AI-assisted news. Scientific Reports, 1–15. https://doi.org/10.1038/s41598-026-58743-0
Summary
This preregistered survey experiment (N=599) with German-speaking Swiss participants investigates whether readers can distinguish, and how they respond to, news excerpts that are human-written, AI-assisted (GPT-4 Turbo rewrites), or fully AI-generated (GPT-4 Turbo from title and lead). The design crucially separates two moments: participants first rated article quality without knowing the source, then were told how each text was produced and asked about engagement intentions. Quality perceptions were statistically equivalent across conditions, but disclosure of AI involvement produced a substantial short-term boost in willingness to continue reading — a boost that did not carry over into stated willingness to read AI-generated news in the future. The authors interpret this pattern as curiosity or novelty-driven engagement rather than genuine acceptance of AI authorship, warning against reading disclosure effects as trust-building.
Key Contributions
- Updated, LLM-era experimental evidence that generative-AI news can match human-written news on perceived credibility, readability, and expertise when authorship is unknown.
- A design that decouples pre-disclosure quality evaluation from post-disclosure engagement intentions, isolating content-level from source-label effects.
- Demonstration that short-term engagement responses to AI disclosure diverge from longer-term reading intentions, refining accounts of algorithmic aversion and source credibility.
- Cross-national evidence from German-speaking Switzerland with a preregistered, openly documented protocol and replication materials.
Methods
Between-subjects online experiment fielded in May 2024 via Bilendi with age/gender quotas. Three conditions used excerpts from SRF articles on Swiss politics: human-written originals, AI-assisted rewrites, and AI-generated texts produced from only the title and lead (10 articles per condition; each participant read two ~150-word excerpts). Quality was measured pre-disclosure on 1–5 Likert scales for expertise, readability, and credibility (Sundar 1999; Haim & Graefe 2017); post-disclosure items captured willingness to continue reading and future willingness to read AI-generated news. Analyses used linear regression with demographic covariates, Benjamini–Hochberg adjustment, Cohen’s d, η² and ω², TOST equivalence tests (SESOI = ±0.20), and article-pair fixed-effects robustness checks.
Findings
- No significant differences across conditions on expertise, readability, or credibility; pairwise Cohen’s d between -0.05 and 0.11; omnibus p > .45.
- TOST tests confirm statistical equivalence within ±0.20 for all three quality dimensions and a composite quality index (α = 0.79).
- Willingness to continue reading was significantly higher in the AI-assisted (b=0.566, d=0.53) and AI-generated (b=0.697, d=0.61) conditions than in the human-written control (both p<.001).
- No significant condition differences in future willingness to read AI-generated news, and equivalence could not be established for this outcome.
- AI-assisted and AI-generated excerpts did not differ significantly from each other on any outcome.
- Results are robust to article-pair fixed effects and to the aggregate quality index.
Connections
This paper speaks directly to work on how audiences perceive and evaluate AI-produced journalism and disclosure labels, connecting to Hameleers2026-mc, Schroeder2026-im, and Dierickx2026-tw on perceptions of AI-generated news, and to Waight2026-ts and Nguyen2026-vm on newsroom and audience-facing dimensions of generative-AI journalism. Its finding that quality perceptions decouple from acceptance also relates to broader debates about AI-mediated persuasion and information exposure explored in Hackenburg2025-dj and DeVerna2025-dl.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify