Label it or leave it? The AI disclosure dilemma

The EU AI Act tells you the minimum you have to disclose, not what your stakeholders will make of you for disclosing it.

Somewhere in your organisation, right now, someone is using AI to write something you'll publish. Maybe the first draft of a press release, or a plain-language summary of a 200-page sustainability report. Perhaps an image in a campaign that no photographer ever shot. The work is already happening. The only open question is whether you say so.

Since 2 August 2026, that question has a legal floor. The EU AI Act's transparency obligations now apply, requiring certain AI-generated content to be marked and disclosed.

But a floor isn't a strategy. The Act only tells you your minimum. It doesn't tell you whether disclosure will make investors trust your reporting more, or candidates trust your employer brand less. It doesn't tell you what happens when a journalist works out that the CEO letter was drafted by a model.

And here's the uncomfortable part: the research points in two directions at once. Audiences say they want disclosure, then they punish the content that carries it.

What you actually have to do

Less than you might assume, and something a little different from what most people expect. The technical heavy lifting isn't yours to do, it’s the companies that build the tools your team uses that have to mark AI-generated output so it can be detected as artificially generated.

Your own duties are narrower, and there are two.

1. You need to disclose so called “deepfakes”, meaning AI-generated or manipulated images, audio, or video resembling real people, places, or events. Worth noting is that deepfake disclosure does not only apply to famous persons. Purely decorative images don’t fall under this category.
2. You need to disclose AI-generated text published to inform the public on matters of public interest, a phrase the Commission reads broadly enough to reach health, financial, employment, and legal communication. The text disclosure duty doesn't apply where the content "has undergone a process of human review or editorial control" and a person or company holds editorial responsibility for publishing it.

Disclosure has a measurable price

The most rigorous evidence here is uncomfortable. Oliver Schilke and Martin Reimann, publishing in Organizational Behavior and Human Decision Processes in 2025, ran 13 experiments across roughly 5,100 participants. Their key finding: "actors who disclose their AI usage are trusted less than those who do not".

The evidence isn't unanimous. A Yahoo and Publicis Media study from February 2024 found that AI disclosures that were noticed lifted ad trustworthiness by 73% and overall company trust by 96%, though that study was commissioned by the advert­ising industry and should be read as such. IAB research from early 2026 found that 73% of Gen Z and Millennial consumers said AI-made advert­ising would make no difference to their purchase intent.

Concealment has a bigger one

The same research contains the finding that could settle the discussion. The trust penalty was worse when AI use was exposed by a third party than when it was disclosed voluntarily. That's where the asymmetry lies. Disclosure is a predictable discount applied at publication, which you can see coming and can offset. Exposure is an uncapped event that converts a conversation about your content into a conversation about your integrity. There are several public backlashes from the past two years that support this.

The questions worth working through

There's no settled answer here, and it would overstate our confidence to pretend otherwise. The EU AI Act sets a baseline, the research pulls in two directions, and practice across the industry, agencies included, is genuinely unresolved. It also isn't only a communications conversation.

Does all content deserve the same treatment?

Almost certainly not. Regulated filings and product claims already carry accuracy duties. Translation and formatting carry little expectation of human authorship. Marketing, employer branding, and anything published in a named person's voice sit in the middle, and that middle is where disagreement happens.

Is a label or an account of the method more useful?

"Made with AI" tells someone very little and may invite the worst assumption. Describing what the AI did and who reviewed it at least shows where accountability sits. What it will not do is protect you. Schilke and Reimann tested six different framings, including one stating that a human had reviewed and revised the work and another that AI was used only for proofreading. All six were trusted less than saying nothing at all. The case for specificity rests on accountability, not on recovering trust, and it's worth being clear internally about which of the two you are buying.

"We don't label work as written by an intern, because we briefed it, researched it, proofread it and edited it. We never said 'researched with Google' either. The standard is the output. If we can't meet it, we shouldn't publish. If we can, I'm not sure we need a label."

Kevin Mullaney, SEO/GEO specialist at Comprend

Are we confusing transparency with quality?

If a reader disregards a text the moment they see "AI-assisted", the label is telling them very little about whether the work is any good.

"A badly researched article is still bad if a human wrote every word of it. A thoughtful, accurate piece doesn't become worse because AI was involved somewhere along the way." 

"If using AI carefully carries the same trust penalty as pressing a button and publishing whatever comes out, we've created a strange incentive. People will simply become less open about using it. That leaves us with less transparency, not more."

Maggie Crusell, Senior Copywriter at Comprend

Is disclosure still a choice?

Until recently this was a decision you controlled. On 11 August 2026 Anthropic said it would watermark text generated by its models to comply with the AI Act's transparency code, with a detection API to follow, and other signatories to the same code are expected to do the same. Light editing probably won't remove it.

"Choosing to tell people how you used AI is one thing. Having the technology mark your use of it is another. Who decides what should be disclosed: the creator, the publisher, the reader, or the company that made the tool?"

Maggie Crusell, Senior Copywriter at Comprend

Does imagery need different treatment from text?

The deepfake duty bites where content could be mistaken for real, which puts photorealistic product renders in a different position from text.

 "Where an AI-generated render could reasonably be read as a photograph of an actual product, a one-line caption saying it's an illustrative render seems right. Not a compliance badge, just a factual note."

Matt Hare, Senior Designer at Comprend

How much does consistency matter?

If marketing labels and HR doesn't, the inconsistency may say more than either choice. Then again, one rule for every function is the fastest route to a policy nobody follows.

Is a published AI policy better than a label?

Webranking's 2025-2026 edition found that just 12% of the 500 largest listed European companies present an ethical AI policy directly on their website. A policy is a different kind of transparency from a label: it describes how you work rather than flagging one piece of content, so it may not trigger the loss of trust that comes with labelling individual pieces.

Who answers for it?

This is the one question we'd argue isn't optional. It's what the EU AI Act now tests, and it's the first thing any stakeholder asks, whether they're a journalist, a regulator, or a candidate who suspects nobody read their application.

Regulation turned out to be the easy part. The EU AI Act draws a narrow line around deepfakes and public-interest text, then hands the rest back with one condition: someone has to be answerable for what gets published.

Everything else is open. Disclosure carries a real but survivable cost, and that cost climbs when someone else reveals the AI use first. Where your line falls is still yours to draw.

That's governance work as much as communication work, and it's the kind of question we help clients think through across reporting, digital channels, and employer brand. If it's on your agenda, we'd be happy to talk.