Disclosure: When to Tell Readers AI Was Involved

Somewhere in your content stack there is a checkbox that says “AI-assisted.” Maybe it’s in your CMS, maybe it’s a line your legal team asked for, maybe it’s a footer your agency added after a nervous client call in 2024. And if you’re honest, it fires on everything: the blog post you outlined in Claude and rewrote by hand, the meta description Jasper spat out, the case study where AI did nothing but clean up your transcript.

That blanket approach feels safe. It isn’t. It’s noise, and noise is worse than silence because it trains readers to skip the one disclosure that actually matters.

The problem with disclosing everything

Run the numbers on a typical five-person content team. Say you ship 12 blog posts, 40 social assets, 8 emails and 3 gated pieces a month. That’s 63 assets. If AI touched 55 of them in some form (spell-check on steroids, a headline variant, a summary for the newsletter), and you tag all 55, you’ve created 55 identical badges saying nothing.

Now the one piece where AI generated a statistical claim you didn’t verify carries the exact same label as the one where Grammarly fixed your commas. A reader who cared cannot tell the difference. You’ve disclosed your way into zero information.

There’s a second cost. Blanket labels get read as a defensive crouch. An “AI-assisted” stamp on a founder interview transcript raises a question that didn’t exist before: what, exactly, did the machine do to this person’s words? You’ve introduced doubt to solve a problem you didn’t have.

The materiality threshold

Here’s the rule I’d hold every asset against. Disclose when AI materially shaped one of three things:

Claims. Did AI generate a fact, figure, statistic, date, quote or causal assertion that a human didn’t independently source? This is the big one. If ChatGPT produced “email marketing delivers £42 for every £1 spent” and nobody traced it back to the DMA report, that claim is AI-originated and the reader deserves to know the provenance is thin. Better still: verify it, cite it, and you’ve eliminated the need to disclose anything.

Voice. Did AI write the prose that a reader will attribute to a named human? A post bylined “by Sarah Mensah, Head of Content” where Claude produced 80% of the sentences is a different object from one where Sarah wrote it and used AI to tighten paragraphs. Bylines are a promise. If the promise is now partly synthetic, say so.

Persona. Did AI generate or simulate a person, a testimonial, a customer quote, a reviewer, a face, a voice? This is where “should we disclose” stops being an editorial judgement and starts being a legal one.

Everything else is tooling. Nobody discloses Grammarly. Nobody discloses that Ahrefs suggested the keyword. Nobody discloses that Notion AI summarised the meeting where the brief was agreed. AI that helps you plan, research, cluster, outline, transcribe, tag, or measure has not shaped what the reader receives as a claim about the world.

Running the threshold on five real assets

The LinkedIn carousel is the one people argue about. My position: a brand account makes no personal authorship claim, so AI-written copy on it doesn’t cross the voice line. The moment your Head of Growth posts it from their personal profile with “here’s what I’ve been thinking about,” it does.

Where UK rules remove your discretion

Editorial judgement gets you most of the way. Then there’s the part where you don’t get a vote.

The CAP Code is the binding text for UK non-broadcast marketing, and rule 2.1 requires that marketing communications be obviously identifiable as such. The ASA has repeatedly applied this to influencer content and to anything where the commercial nature is disguised. The relevant application to AI is narrower than people assume: the ASA isn’t policing whether Claude wrote your blog post. It’s policing misleadingness (Section 3) and testimonials (Section 3.45 onwards), which require that testimonials be genuine, documented and held on file.

So: a synthetic customer quote is not a disclosure problem, it’s a breach. An AI-generated “before and after” is a breach. An AI-generated image of a product benefit that the product doesn’t deliver falls under 3.1 and 3.11 regardless of how you label it. You cannot disclose your way out of a misleading claim, and this is the single most common misunderstanding I see in governance docs.

The Digital Markets, Competition and Consumers Act 2024 came into force in April 2025 and moved unfair commercial practices enforcement to the CMA, with direct fining powers of up to 10% of global turnover. Fake reviews are now explicitly banned, including commissioning, publishing or hosting them. If your review-generation workflow includes a step where an LLM drafts review text, stop the workflow.

Platform rules are stricter and more specific than advertising law:

• Meta requires disclosure of AI-generated or digitally altered photorealistic video, realistic audio, and images where content is generated by AI in a way that risks deceiving the public; it applies “AI info” labels automatically via C2PA and IPTC metadata, and requires self-disclosure where detection fails.
• TikTok mandates the AI-generated content label for realistic synthetic media and auto-applies it through Content Credentials. Undisclosed realistic AI content is removable.
• YouTube requires creators to declare “altered or synthetic” realistic content in Studio at upload, with a label in the description or on the player for sensitive topics.
• Google Search doesn’t require disclosure and has said so plainly, but its spam policies target scaled content abuse: content produced at volume primarily to manipulate rankings, whether or not AI made it.

Note the pattern. Every mandatory rule keys on realistic synthetic depiction of people or events, not on whether a language model helped you write. That’s the same materiality logic arriving from the regulatory side.

The EU AI Act matters for UK teams with EU audiences: Article 50 transparency obligations apply from 2 August 2026, covering deep fakes and AI-generated text published to inform the public on matters of public interest. Most B2B content marketing sits outside that, but “text published to inform the public” is doing a lot of work in that sentence and you should watch how it’s interpreted.

What good disclosure actually says

When you do cross the threshold, the standard footer fails. “This article was created with AI assistance” tells a reader nothing about what to distrust.

Write disclosures that scope the involvement:

The five industry benchmarks in this piece were drafted using Claude and verified against the original sources, which are linked. The analysis and recommendations are the author’s.

Or:

The presenter in this video is a synthetic avatar (HeyGen). The script was written by our product team.

Both take fifteen seconds to write and give a reader something to act on. Place them where the material thing happens: next to the stats block, under the video, in the byline area if voice is the issue. A footer disclosure for a claim in paragraph three is technically true and practically useless.

Building the check into the workflow

Make it a field, not a policy document. Three questions in your brief template, answered before publish:

Did AI generate any claim, figure or quote not verified against a source?   [ ] Y  [ ] N
Is this bylined to a named person whose prose AI substantially wrote?       [ ] Y  [ ] N
Does this contain synthetic depiction of a person, voice or likeness?       [ ] Y  [ ] N

Any Y triggers a scoped disclosure. Three Ns and you ship clean. In Asana or Notion this is a three-checkbox property on the content item; in a Google Sheet tracker it’s three columns. The point is that it’s answered per asset by the person who made it, at the moment they’d remember.

Log the answers. When a client or your legal counsel asks in eight months what your AI policy was in September, a spreadsheet showing 63 assets assessed and 4 disclosed is a far better answer than a policy PDF nobody followed. That audit trail is also the backbone of the wider control set covered in governance, disclosure and UK compliance, which is where things like model approval, data handling and client sign-off belong.

The bit that stings

The materiality threshold has an uncomfortable implication: most of the time, the honest answer to “should we disclose this?” is “no, but you should have verified that stat.”

Disclosure is often a substitute for doing the work. Labelling a piece “AI-assisted” because you’re not sure the numbers hold up is not transparency, it’s insurance. The reader can’t check either. If a claim is shaky, fix the claim.

The teams I’ve seen handle this well disclose less than they used to and verify more. They killed the blanket checkbox, added the three questions, and found that their disclosure rate dropped from 100% to something like 6%. What went up was the number of stats with a source link next to them, which is the outcome anyone asking for transparency actually wanted.

Go and look at your last twenty published assets. Run the three questions. If the answer comes back N, N, N on all twenty and you disclosed on all twenty, you have a labelling habit, not a disclosure policy. Start by deleting the footer.