Study reveals AI news disclosures backfire—readers want detail-on-demand

A new study by Pooja Prajod, accepted to the CHIWORK Workshop on Interrogating GenAI Augmentation, challenges how newsrooms communicate AI involvement to readers. The paper, available on arXiv, analyzes two current disclosure approaches: brief one-line labels and detailed disclosures specifying human oversight and error reporting. A controlled experiment with 34 news readers found that neither works as intended. Detailed disclosures triggered what the author calls a 'transparency dilemma'—instead of increasing trust, they actively reduced it. One-line labels avoided that pitfall but created an information gap, prompting readers to expend cognitive effort searching for signs of AI involvement that the disclosure hinted at but didn't explain.

Crucially, readers aren't rejecting transparency itself. When asked to design their ideal system, they proposed four concrete features: detail-on-demand (click to expand), proportional AI-ratio visualizations (e.g., a slider showing human vs. AI contribution), outlet-level signals (a standard badge across articles), and explicit 'no AI' labels for fully human content. The author argues the disconnect between what journalists believe is responsible disclosure and what users actually need is fundamentally a design problem—one the HCI community must address. The study suggests that static, journalist-crafted disclosures fail because they don't account for reader agency or cognitive load, and that interactive, user-controlled systems could restore trust in an era of generative AI in newsrooms.