Skip to content
Article Nielsen Norman Group Sep 2026

NN/g: the PACED framework for deciding when to disclose AI use

One of the more awkward questions in AI product design is deceptively simple: should you tell users when AI generated or influenced something? Research on this question has produced contradictory results, and Taylor Dykes’s September 2026 article for Nielsen Norman Group addresses why that is — and offers a structured way to make the decision.

The disclosure paradox

The intuitive assumption is that transparency builds trust and that disclosing AI use is therefore always the right move. The evidence does not support this cleanly. In some product contexts, users who knew AI was involved reported higher satisfaction; in others, the same disclosure reduced trust or made the interaction feel inappropriate. Dykes frames this as the disclosure paradox: the effect of disclosure depends on factors that vary by audience, context, and the nature of the AI involvement itself.

The PACED framework

Dykes proposes a five-factor framework — PACED — to help teams decide whether to disclose AI involvement and how to frame it.

Policy is the starting point. Many industries now have regulatory or company-level requirements for AI disclosure. Before applying any other factor, designers need to know what the floor is. Where policy mandates disclosure, the decision is already made.

Audience asks who the users are and what relationship they have with AI generally. Users who work with AI tools daily bring different priors than those encountering AI-assisted content for the first time. The same disclosure that reassures one group can alarm another.

Context concerns the nature of the task. High-stakes contexts — healthcare recommendations, legal summaries, financial guidance — carry different disclosure obligations than low-stakes ones like a playlist suggestion or an auto-generated caption. Users tend to apply more scrutiny to consequential outputs, and disclosure signals that scrutiny is warranted.

Expectations looks at what users assume about the product already. If users broadly expect human-authored content in a given interaction, learning otherwise creates a gap that affects satisfaction. If the product category has normalized AI involvement, disclosure adds less friction.

Degree covers the actual level of AI contribution. A fully generated response differs from one where AI drafted and a human edited, which differs again from AI-assisted routing that the user never sees. The article argues for disclosing in proportion to degree: blanket “AI-powered” labels tell users little and can read as noise, while specific, contextualized disclosure — “this summary was generated from the transcript” — gives users something actionable.

Why it is useful

The framework does not produce a binary answer so much as a structured argument. Teams can use it in design reviews to test whether a disclosure decision holds across the five factors, and to surface disagreements about audience assumptions or context classification before they reach production. It is most directly applicable to product designers, content designers, and legal or policy reviewers working on AI-assisted features where the disclosure question is live.

The article assumes no technical background and is short enough to read in a single sitting.