Disclosing AI: Contrasting Effects and the Paradox of Transparency

No rules are imposed for signaling the use of AI in content. Variable reactions are observed depending on the context, the extent of the tool's intervention, and the target audience. Several studies describe a paradox: readers say they want transparency while rating content that displays it more harshly. A framework like PACED is proposed to guide the reflection.

Penalties Observed Despite a Demand for Transparency

When a human intervention, sincerity, or relationship is expected in communication, the disclosure of AI usage often accompanies negative evaluations. Research indicates that this type of transparency can make the content or its author appear less authentic. In a comprehensive study conducted by Jasper David Brüns and Martin Meißner, indicating that a fashion image integrated into a brand's social posts was generated by AI led to lower ratings on brand authenticity, publication credibility, and attitude toward the brand. This finding aligns with what Siavosh Sahebi, Paul Formosa, and Sarah Bankins call the "disclosure paradox": readers claim they want AI usage to be disclosed but react negatively when it is. In their study, a text attributed to AI is judged as less reliable, less authentic, and less useful, even as participants affirm the importance of disclosure.

No Uniform Effect Measured in the Literature

The available results do not describe a systematic negative reaction to transparency regarding AI. The reviewed experiments and reports do not yield a uniform public reaction: depending on the contexts, disclosure sometimes leads to judging content as less reliable, authentic, or qualitative, and sometimes has no measurable effect. After analyzing 47 studies on how readers react to authority and disclosure in information, Lorena Licenji and Julian Hoxha do not identify a homogeneous trend. These observed differences can be explained by the type of news, the perception of the publisher's credibility, or the specifics of the audience. Among all this research, only 10 studies explicitly examined the impact of disclosure; they too do not find a systematic influence on trust or credibility. In this context, no single rule can decisively address the question of whether to disclose or not; a framework like PACED is proposed to assist in decision-making.

The Role and Extent of AI Affect Evaluation

Beyond simply knowing that an AI was involved, understanding its role and contribution weighs on the reaction. In the experiment by Zhuoyan Li and colleagues, announcing that AI generated the content lowers quality evaluations for argumentative essays as well as for creative stories. In contrast, specifying that the AI only edited the text has a limited effect: the decline appears for creative stories, but not for essays. Hiroki Nakano and his team find a similar pattern. Readers first evaluate an author, then learn that a given percentage of the text was generated or edited by AI before re-evaluating; their judgments decrease as the communicated percentage of involvement increases.

Attitudes, Self-Confidence, and Familiarity Modulate Trust

Audience characteristics also play a role. By aggregating 12 out of 13 experiments, Oliver Schilke and Martin Reimann find that the trust penalty associated with disclosure is lower among individuals with a more favorable attitude toward technology and those who perceive AI as accurate, though it does not disappear entirely. Readers' self-reported confidence in their writing abilities also influences reactions: in the work of Zhuoyan Li et al., those who see themselves as more confident writers lower their evaluations more after disclosure, while the less confident tend to be less negative overall. Familiarity with AI comes into play, but without consistency: being more familiar with ChatGPT sometimes makes disclosure more acceptable, and sometimes not.

What Disclosure Covers and What It Is Not

An AI disclosure is a human-readable mention that explains how the tool contributed to the content. It can take various forms: an icon attached to a caption on a social network, a footer note, or a phrase integrated into a long text. It can be added proactively by the creator or provided by a third party such as the hosting platform; examples include a warning at the end of an article at KXAN and a mention applied in Google Docs. Disclosures are distinct from provenance marks and watermarks, such as the statistical watermark published by Anthropic for its Claude models: the latter are intended for machines and are not detectable by the human eye or a screen reader alone, whereas a disclosure must remain accessible without additional tools. In a preliminary study in fashion imaging, participants did not reliably differentiate AI-generated visuals from those created by humans, highlighting the importance of explicit information when it is chosen.