AI Disclosure Statement Examples for UK Marketing Teams

Most disclosure statements fail for the same reason: they were written once, by someone in legal, to cover every conceivable use of AI, and so they say nothing. “This content may have been created with the assistance of artificial intelligence” tells a reader precisely as much as “this content may contain words.” It protects nobody, informs nobody, and signals that you haven’t thought about it.

A useful disclosure does three things. It names what the AI actually did, it names who is accountable for the result, and it does that in fewer than 40 words so it can sit at the bottom of a blog post without becoming the blog post. What follows is a set of statements you can lift and adapt, organised by how deeply AI was involved, plus the placement and process decisions that make them defensible. For the wider regulatory picture and how disclosure fits alongside data handling, attribution and sign-off, see the governance, disclosure and UK compliance pillar.

What UK rules actually require right now

There is no UK statute that says “label your AI-written blog posts.” The government’s approach has been to let existing regulators apply existing law, which means your obligations come from three directions.

The CAP Code governs anything that qualifies as a marketing communication. Rule 3.1 says ads must not materially mislead. Rule 3.7 says you must hold documentary evidence for objective claims before publication. Rule 2.1 says marketing communications must be obviously identifiable as such. None of these mention AI, and all of them bite harder when AI is in the loop, because a model will produce a confident statistic with no source behind it and your substantiation file will be empty.

The DMCC Act 2024 brought the unfair commercial practices regime into force on 6 April 2025, with fake reviews and fake endorsements sitting on the banned-practices list. Generating a testimonial, a customer quote, or a “review” with a model is now squarely a consumer protection problem, not a taste problem.

Then there’s the EU AI Act, which matters if you publish to EU audiences. Article 50 transparency duties apply from 2 August 2026. The text obligation is narrower than most summaries suggest: it covers AI-generated text published to inform the public on matters of public interest, and it falls away where the content has been through human review and a person or organisation holds editorial responsibility for it. Your product comparison page is unlikely to be caught. Your commentary on pension reform might be. Either way, a documented human review step is the thing that resolves it.

Google, for its part, has been consistent that production method isn’t the issue and ranking doesn’t hinge on disclosure. That is a reason to disclose for readers rather than for crawlers, and it changes where you put the statement.

Three tiers, three different statements

Sort every piece you publish into one of three buckets before you write anything. The bucket determines the statement.

Tier 1, AI-assisted. The model helped with research synthesis, outlining, headline variants, meta descriptions, transcript cleanup, or line editing. A human wrote the prose. Most teams land here for 60-70% of output.

Tier 2, AI-drafted, human-rewritten. The model produced a first draft from a brief. A human restructured it, replaced the examples, added the original reporting, and rewrote enough that the published version shares an outline with the draft but not much else.

Tier 3, AI-generated, human-reviewed. The model produced the published text. A human checked facts, tone and claims, and signed off. Programmatic pages, glossary entries, localised variants and product feed descriptions usually sit here.

Anything involving a real person’s quote, a customer story, a named byline, or a factual claim about performance needs the byline question answered separately, and I’ll come to that.

Eight statements you can adapt

Tier 1, standard blog footer (24 words):

Written by Priya Raman. Claude was used for research synthesis and headline testing. All reporting, analysis and final copy are Priya’s.

Tier 1, shorter inline variant for short-form posts (16 words):

Research and outlining assisted by AI. Written, fact-checked and edited by the Ridgeline content team.

Tier 2, long-form article (38 words):

This article was drafted with GPT-5 from a brief prepared by our team, then rewritten, restructured and fact-checked by Tom Beddoes. The benchmark figures in section three come from our own client data, verified October 2026.

Tier 3, programmatic or templated page (29 words):

This page was generated using AI from our verified product database and reviewed by a member of our team before publication. Spot an error? Email [email protected].

Customer case study, where AI touched the write-up but not the words of the customer:

Quotes in this case study are verbatim from a recorded interview with Jo Fairhurst on 14 May 2026 and were approved by her before publication. The surrounding narrative was drafted with AI assistance and edited by our team.

That one is worth more than the others combined. It draws the line exactly where the DMCC draws it.

Gated report or whitepaper, front matter:

Methodology note: survey data was collected via Qualtrics from 412 UK respondents between 3 and 21 March 2026. Analysis was performed in Python. AI tools were used for chart annotation and to draft descriptive summaries of the findings; all interpretation, conclusions and recommendations are the authors’.

Synthetic voice or video:

The narration in this video uses a synthetic voice generated with ElevenLabs from a script written by our team. No real person’s voice has been cloned.

Email newsletter sign-off (19 words):

Some sections of this newsletter were drafted with AI. Everything you read was edited and approved by a human first.

Placement matters more than wording

Put Tier 1 and Tier 2 statements at the end of the article, in the same block as the author bio, at 85-90% of body text size. Not in a modal. Not in a tooltip. Not, please, in the cookie banner.

Tier 3 disclosure goes at the top, above the fold. If a reader is about to make a decision based on machine-generated text, telling them afterwards is theatre. On programmatic pages this usually means a single line under the H1 in muted text, around 120-160 characters.

Video and audio disclosure belongs in the first 10 seconds or in a pinned on-screen card, because YouTube descriptions get read by roughly nobody. YouTube’s own altered-content toggle handles the platform-level label; your spoken or on-screen line handles the human one.

The policy page behind the footnote

Every footer statement should link to a single URL, something like /ai-policy or /how-we-use-ai. That page carries the detail the footnote can’t: which tools you use by name, what data never goes into them, whether you use AI for customer-facing chat, how you handle corrections, and who to contact. Keep it to 500-700 words and date-stamp it. Review it whenever you add a tool to the stack, which for most teams is more often than quarterly.

One paragraph that page needs and most omit: what you don’t do. “We do not generate customer quotes, reviews, testimonials or case study narratives without the named individual’s review and approval. We do not use AI to produce images of people who do not exist and present them as customers or staff.” Specific negatives read as credible in a way that positive assurances never manage.

Bylines, and the thing nobody wants to decide

If a piece is Tier 3 and carries a named human byline, you have a problem no footnote fixes. The byline is the claim. Options, in order of how well they survive scrutiny: byline it to the brand (“Ridgeline Editorial”), byline it to the person who did the review with the role stated (“Reviewed by Sam Okafor, Head of Content”), or drop the byline entirely for that content type. Agency leads should settle this in the SOW rather than at publication, because the client’s brand is the one carrying the risk and their tolerance won’t match yours.

Resist the urge to run everything through Originality.ai or GPTZero as a gate. OpenAI withdrew its own AI text classifier in July 2023 after it correctly flagged only 26% of AI-written text while misflagging 9% of human writing, and the independent tools have not solved the problem since. A detector score is not evidence, and using one to police your own freelancers will cost you good writers.

Machine-readable signals

For images, IPTC’s digitalSourceType field with the value trainedAlgorithmicMedia is the established marker, and Google Images reads it. Adobe’s Content Credentials write C2PA provenance data from Photoshop and Firefly. For text there is no equivalent standard with meaningful adoption, and schema.org has nothing purpose-built. Don’t invent a custom property and expect anything to consume it.

What usually goes wrong

Teams write the statement, publish it, and then quietly change their workflow. The footer still says “research assistance” six months after the team moved to full drafting in Jasper. That drift is the actual compliance exposure, not the wording.

Set a calendar item for the first working week of each quarter. Pull ten published URLs at random, ask the person who made each one which tier it was, and compare that to what the page says. Where the two disagree, you’ve found either a stale statement or a process nobody told you about.