When should you disclose AI use in written content?

Taylor Dykes, writing for Nielsen Norman Group, argues that "no single rule determines whether you should disclose" AI use in text content, because the research reviewed for the article shows audience reactions to AI disclosure "varied across studies and contexts": sometimes disclosure lowered ratings of trustworthiness, authenticity, or quality, and sometimes it had no measurable effect. In place of a universal disclose-or-don't rule, Dykes proposes working through "the five PACED factors — policy, audience, context, expectations, and degree of AI contribution — as a set of questions to decide whether to disclose."

Each factor names a different source of variation the article found in the literature. Policy covers legal and platform rules, such as the EU AI Act's transparency obligations (in force for particular cases since August 2026), plus the observation that "third-party AI disclosure triggered the strongest negative reactions," which argues for disclosing proactively before a platform labels the content instead. Audience covers a reader's preexisting attitude toward technology: people favorable toward AI "may react less negatively to disclosures," while writers "confident in their own writing may react more negatively" to disclosed assistance. Context covers how much the content's value rests on qualities "people normally associate with a human author — sincerity, creativity, or personal connection," so fiction and interpersonal writing draw more scrutiny than utilitarian text. Expectations and degree of AI contribution are named but left less specified in the excerpt; they extend the same logic to how much human judgment a reader expects and how much of the content the AI actually produced versus merely edited.

The framework's audience factor leans on the same asymmetry documented in Do readers and writers differ on AI disclosure necessity? and Why do readers and writers disagree on disclosure necessity?, both built on Fang, Wen and Lee's finding that readers want disclosure more than writers want to give it. Its policy factor matches Do readers trust unlabeled AI-written messages as much as human ones?, where a disclosed origin, not AI use itself, drove the penalty. And its premise that effects are contingent on context reframes the flatter penalties in Does disclosing AI assistance make readers trust articles less? and the broad 13-experiment penalty in Does disclosing AI use damage how trustworthy you seem? as data points inside a contingent system rather than a general law. It also names, as a reason no single rule applies, the gap described in Why do people penalize AI disclosure even when they request it?.

The excerpt names "expectations" and "degree of AI contribution" as two of the five factors but does not define them with the same specificity as policy, audience, and context, so PACED reads as a checklist derived from a literature review rather than a tested decision procedure — Dykes does not report having validated it against outcomes. The implication, held at the strength the evidence allows, is that disclosure guidance for AI-generated text should be built as a situational checklist rather than a blanket rule, until someone tests whether weighing the five factors actually predicts audience reaction better than disclosing, or not, by default.