When should a brand disclose AI use? Research says it depends

Should you disclose AI use when it helped make your content or product? The research answer is: it depends on who is reading, what kind of content it is and how much AI did. On September 25, 2026, Nielsen Norman Group published the PACED framework, five questions that help teams decide when to disclose AI use, and in the EU some disclosures have been a legal requirement since August 2 ^{[1, 2]}.

We get this question from clients almost weekly, about blog posts, product descriptions, chatbots and campaign images. Here is what the research says and how we turn it into a policy.

The disclosure paradox

The NN/g article, written by Taylor Dykes, starts from a contradiction the research keeps finding. People say they want AI use disclosed, yet they react negatively to content that carries a disclosure ^[1]. They also want others to disclose their AI use more than they want to disclose their own ^[1].

The effect is not uniform. NN/g cites a review of 47 studies on AI disclosure in news that found no consistent trust or credibility penalty ^[1]. Other work found that disclosing that AI generated an essay or story lowered quality ratings, while disclosing that AI only edited it had a narrower effect: ratings fell for creative stories but not for essays ^[1].

One finding stands out for brands. In research by Schilke and Reimann, people trusted a company that disclosed AI use less than one that said nothing, but they trusted a company whose AI use was exposed by a third party least of all ^[1]. Hiding AI use only works until someone notices.

The PACED framework, factor by factor

PACED turns these findings into five questions ^[1]:

• Policy: Do laws or platform rules require disclosure? NN/g points to the EU AI Act and to platforms that apply labels automatically ^[1].
• Audience: How does this audience feel about technology and AI? People with more favourable attitudes showed smaller trust penalties ^[1].
• Context: What kind of content is it? Readers judge personal and interpersonal writing more harshly than transactional messages ^[1].
• Expectations: Does the reader assume a human made this, or assume they would be told otherwise? ^[1]
• Degree of AI use: Did AI generate the content, or help edit it? The more AI contributed, the stronger the case for disclosure ^[1].

NN/g also notes that readers are most likely to penalise AI use when they believe it replaced a meaningful human contribution, such as creativity, sincerity or personal judgment ^[1]. A thank-you note from a founder is a very different case from a shipping notification.

What the EU AI Act now requires

Policy is the one factor that is not a judgment call. Article 50 of the EU AI Act has applied since August 2, 2026 ^[2]. According to the European Commission, it requires that:

• People are told they are interacting with an AI system from the start of the first interaction, unless it is obvious ^[2].
• Generative AI providers mark synthetic audio, images, video and text in a machine-readable way ^[2].
• Deepfakes are clearly labelled, with lighter, appropriate disclosure for artistic or satirical work ^[2].
• AI-generated text on matters of public interest is labelled unless it went through human review or editorial control, and spell-checking alone does not count as review ^[2].

The Commission lists fines of up to €15 million or 3% of worldwide annual turnover, and a grace period until December 2, 2026 for marking obligations on systems already on the market ^[2]. Indonesian brands that sell into Europe, or run chatbots for EU users, are in scope for the parts that apply to them.

When to disclose AI use: how we apply PACED

In practice, three rules cover most cases. First, any interface where a user talks to an AI, such as a support assistant, says so up front, which matches both the EU rule and user expectations ^[2]. Second, content where AI did the heavy lifting and the reader expects a human voice, such as founder letters or reviews, either gets a real human rewrite or a clear disclosure. Third, routine content where AI helped with grammar or formatting does not need a label, but the team keeps a record of how it was made.

The last point is easy to skip. If a journalist or competitor asks later, you want an answer ready, because third-party exposure is the scenario that damages trust most ^[1].

Key takeaways

• Reactions to AI disclosure vary with audience, context and how much AI was used ^[1].
• Being exposed by someone else damages trust more than disclosing up front ^[1].
• NN/g’s PACED framework covers Policy, Audience, Context, Expectations and Degree of AI use ^[1].
• EU AI Act Article 50 transparency rules have applied since August 2, 2026 ^[2].
• A short written disclosure policy prevents inconsistent, last-minute decisions.

Quick answers

Do we have to label every AI-assisted blog post? Not by default. In the EU, text on public-interest matters needs a label unless it had real human review or editorial control ^[2]. Elsewhere, use the PACED questions.

Does disclosure hurt conversion? It can lower ratings in some contexts, especially creative or personal content ^[1]. Test wording with your own audience.

What about AI-generated images in ads? Label anything that could pass as a real person or event, and check platform rules, which may label it for you ^{[1, 2]}.