Technical Summary: Designed by Journalists, but Is It for Readers? Rethinking AI Disclosures and Transparency in News

Problem Statement

As newsrooms increasingly integrate generative AI (GenAI) for drafting, summarizing, and translating content, a critical accountability gap has emerged regarding how to disclose AI involvement to readers. Current disclosure practices generally fall into two categories: brief one-line labels or detailed statements specifying human oversight, editorial accountability, and error-reporting mechanisms.

The core problem is a disconnect between practitioner assumptions and user experience. Journalists and ethics frameworks operate under the assumption that "more transparency equals more trust," leading to detailed disclosures. However, this approach fails to achieve its goal. Furthermore, while the EU AI Act exempts AI-assisted text with human editorial oversight from mandatory disclosure, many outlets disclose voluntarily. The paper argues that current disclosure designs are treated as compliance checkboxes rather than user-centered design challenges, potentially creating "dark patterns" where disclosures exist technically but are ignored or misunderstood by readers.

Methodology

The paper synthesizes findings from a controlled experiment involving 34 news readers (cited as [12] in the text) and qualitative insights from the same study.

• Experimental Design: Participants were exposed to news articles with varying disclosure formats:
  1. No disclosure.
  2. One-line disclosure: A brief label (e.g., "this article was edited using AI").
  3. Detailed disclosure: Comprehensive statements including human oversight, responsible editors, and error-reporting channels.
• Analysis: The study measured reader trust levels and cognitive engagement. It also analyzed participant feedback regarding their ideal disclosure designs, moving beyond testing existing norms to gathering user-generated concepts for future design.

Key Findings and Results

1. The Transparency Dilemma

The controlled experiment revealed a counter-intuitive result: detailed disclosures reduced trust compared to simple one-line disclosures.

• Detailed Disclosures: Triggered a "transparency dilemma," where the additional context paradoxically lowered reader trust rather than building it.
• One-Line Disclosures: Did not reduce trust compared to no disclosure but introduced an information gap. Readers aware of AI involvement but lacking details expended extra cognitive effort scanning the text for signs of AI, creating friction without providing clarity.

2. The Risk of Dark Patterns

The paper posits that if detailed disclosures become standard but are routinely scrolled past or ignored, they function as dark patterns. They serve the outlet's compliance needs and create an "illusion of transparency" without actually informing the reader, potentially eroding trust further.

3. User-Generated Design Concepts

Contrary to the assumption that readers want less information, participants proposed designs that prioritize user agency. They did not reject transparency; they rejected imposed transparency. The study identified six specific design concepts suggested by readers:

• Highlight-for-glancing: Visual highlighting of AI-involved steps, allowing for quick scanning with optional deep-dive details.
• Info Button: A brief label with an interactive element (e.g., "i" icon) allowing users to access detailed disclosures on demand.
• Outlet-level Disclosure: A centralized statement at the organization level regarding AI policy, reducing repetitive per-article fatigue.
• Proportional AI-ratio Visualization: Visual indicators (e.g., percentage bars or partially colored pages) showing the proportion of AI contribution.
• Visual Trust Stamp: A recognizable badge indicating responsible AI use, analogous to organic food labels.
• "No AI Used" Label: An explicit negative label to signal the absence of AI, reframing disclosure as a bidirectional trust signal.

4. Regulatory Expectations

Despite the EU AI Act's exemption for text with human oversight, 71% of participants (24/34) argued that AI disclosure in news should be government-regulated. However, the paper emphasizes that regulation must focus on the form of disclosure, not just its existence, to avoid the backfire effects observed in the experiment.

Key Contributions

1. Empirical Evidence of the Transparency Dilemma: The paper provides data showing that the standard industry practice of detailed AI disclosures can actively harm trust, challenging the prevailing "more is better" heuristic in AI ethics.
2. Reframing Disclosure as an HCI Design Problem: It argues that the failure of current disclosures is not an ethical failure but a design failure. The gap between what journalists think is responsible and what users find trustworthy requires a shift from compliance-driven to user-centered design.
3. A Toolkit for User-Centric Disclosure: The paper visualizes and categorizes six reader-proposed design patterns that shift control to the user (detail-on-demand, proportional visualization, etc.), offering a concrete alternative to static text blocks.
4. Cross-Domain Reflection: The author extends these findings to the broader HCI and research community, questioning whether ACM's AI use statements in papers suffer from the same "transparency dilemma" and whether researchers design for their users or merely for their own professional instincts.

Significance and Claims

The paper's primary significance lies in its challenge to the assumption of linear trust-building through transparency. It claims that:

• Current AI disclosure practices in journalism are failing because they are designed for practitioners (compliance) rather than readers (trust).
• The "transparency dilemma" suggests that mandatory, detailed disclosures may be counterproductive if not designed with user agency in mind.
• The solution lies in interactive, proportional, and on-demand disclosure designs that allow readers to control their engagement with transparency information.
• This disconnect is a systemic issue in the HCI community, where disclosure norms are often institutionalized based on professional intuition rather than empirical evidence of user reception.

The paper concludes that until disclosure designs are tested against the actual needs and behaviors of the end-user, they risk becoming performative compliance tools that undermine the very trust they aim to build.