The ASA has no blanket AI disclosure rule
The UK test is whether omitting AI involvement misleads, not whether AI was used. The cases that do require disclosure, and what the advertiser still carries.
There is no rule in the UK that says an advertiser must announce that an ad was made with AI. Teams keep looking for one, keep failing to find it, and keep concluding they have missed something. They have not. The ASA's position, set out in its own guidance on disclosure of AI in advertising, is that there is no blanket duty to disclose AI use.
That sounds permissive until you read the test that replaces it, which is harder to satisfy than a label would be.
General information, not legal advice. Ad clearance for a specific campaign needs specialist input.
The test is misleadingness, not method
The CAP Code does not ask how an asset was produced. It asks whether the ad, taken as a whole, is likely to mislead. AI involvement becomes disclosable when hiding it is what does the misleading, and stays undisclosable-by-default when it does not.
This is a genuinely different question from the one most disclosure policies are built around. A policy that asks "did we use a generative model?" produces a label on every asset, including the ones where a label adds nothing and the ones where a label is actively unhelpful. The UK question is "does the audience end up with a false impression because we did not say?" Those two questions disagree constantly.
A stock-style background plate generated rather than licensed does not change what a viewer believes about the product. A generated image of the product doing something the product cannot do changes it completely. Same tool, same pipeline, opposite answers.
The cases that do require disclosure
The ASA's guidance identifies the shapes where omission misleads. These are the ones to build your review around.
AI-generated imagery implying real product performance. This is the big one and it catches more work than teams expect. A generated hero shot showing a result the product does not achieve is a misleading ad whether or not the words claim it. Before and after treatments, food and drink, cosmetics, home improvement, anything where the image is doing the claiming — the image carries the same evidential burden as the copy.
Synthetic endorsements. A generated person recommending a product reads as a customer or an expert. If no such customer or expert exists, the endorsement is fabricated, and the fact that it was rendered rather than filmed is exactly the material fact that has been withheld.
Unachievable results. Related to the first but broader. Where the visual sets an expectation the ordinary user will not meet, the visual is the claim, and "it's illustrative" is not a defence that survives contact with the Code.
Notice what these have in common. In none of them is the problem "AI was used." The problem is a false impression, and AI is simply the mechanism that produced it cheaply. The same ads would breach if you had achieved the same false impression with a retoucher and a paid actor.
The responsibility that does not move
The part of the ASA's position that gets least attention is the part with the most operational weight: advertisers are fully responsible for AI-produced content.
There is no diminished-responsibility tier for output you did not personally compose. Not for a model's hallucinated claim in a script. Not for a generated statistic. Not for a comparison the tool invented while filling out a template. If it ran in your ad, it is your ad.
That has three practical consequences.
- Substantiation applies to generated claims exactly as it does to written ones. Every factual claim in the finished asset needs evidence held before it runs, regardless of which part of the pipeline drafted it. A model producing a plausible-sounding percentage is producing an unsubstantiated claim, and plausibility is not substantiation.
- "The model produced it" is not a mitigating circumstance. It is a description of your process, and your process is your responsibility.
- Review has to happen on the finished asset, not the brief. Generated detail arrives after the brief is signed off. If your approval gate is upstream of generation, it is inspecting something that no longer exists by the time the ad ships.
The governance layer for AI content is where this gets solved structurally, and our walkthrough of writing an AI disclosure line nobody scrolls past covers the wording once you have decided a line is needed.
The UK answer does not travel
This is where teams get caught. They read the ASA position, correctly conclude that no blanket label is required, and then apply that conclusion to a campaign that also runs in the EU and New York.
| Market | Position |
|---|---|
| UK | No blanket disclosure duty. Test is whether omission misleads |
| EU | AI Act Article 50 applies from 2 August 2026. Deployers must label deepfakes, disclosed to the viewer on first exposure rather than only in metadata |
| New York | Conspicuous disclosure required when an ad reaching NY consumers features a fully synthetic human performer, effective 9 June 2026 |
| US federal | No AI-specific advertising rule. Section 5 deception and unfairness apply |
The EU rule and the New York rule are duties triggered by the content type, not by whether omission would mislead. An EU deepfake gets labelled whether or not anyone would have been misled. A New York ad with a fully synthetic performer gets a conspicuous disclosure on the same basis, with penalties set at $1,000 for a first violation and $5,000 thereafter.
Article 50 carries a carve-out for evidently artistic, creative, satirical or fictional works, which need only a disclosure that does not hamper enjoyment of the work. New York's statute has an exemption covering promotion of expressive works. Neither carve-out is a general advertising exemption, and reading them as one is a common and expensive mistake. Our breakdown of Article 50 and sponsored content goes through where the deployer duty actually lands.
The result is that a UK-only campaign and a multi-market campaign have genuinely different answers, and the multi-market answer is the stricter one applied across the board unless you are willing to maintain divergent cuts. Most teams are not, which is why the cross-platform labelling checklist tends to collapse into a single house standard.
A working rule
The version that survives contact with a real review cycle:
- Ask what the audience will believe about the product, the person and the result after seeing the asset. Not what it was made with.
- If any of those beliefs would be false without the disclosure, disclose. That is the UK duty, and it is a duty even though it is not a blanket one.
- If the asset ships to the EU or New York, apply their content-triggered duties on top. They do not care about your misleadingness analysis.
- Substantiate everything the finished asset claims, including claims that arrived from the model rather than the copywriter.
- Check platform policy separately. Platform disclosure requirements are contractual, stricter than law in places, and change without notice. They sit outside all of the above.
Where teams land after running this a few times is a standing AI disclosure policy that answers the question once rather than per-asset, plus a shared understanding of what an AI content label actually communicates versus what people assume it does. If you are building a recurring synthetic presenter, that policy is where the standing answer belongs.
FAQ
Does the ASA position mean I can never be pulled up for undisclosed AI?
No. It means AI use is not itself the breach. If omitting AI involvement is what creates a false impression — a synthetic endorsement, a generated demonstration of performance — that is a breach, and the absence of a blanket rule does nothing to help.
Is a generated background different from a generated product shot?
Under the misleadingness test, usually yes. A background plate that changes nothing a viewer believes about the product is a production choice. A product shot is a claim about the product. Same tool, different function in the ad, different answer.
If I label everything, am I safe in the UK?
Safe from this particular question, at some cost. Over-labelling does not breach the Code, but it also does not cure a misleading claim. An ad that shows an unachievable result is still misleading with an AI label on it, because the label discloses the method rather than correcting the impression. Labels do not substitute for substantiation. See our overview of synthetic media disclosure for what a label does and does not carry.
What about UK ads that only run on social platforms?
Both regimes apply. The Code applies to the ad; the platform's own synthetic media policy applies as a condition of distribution. Platform rules are frequently stricter and are enforced by automated classifiers that will not be reading your misleadingness analysis. Our AI ad disclosure and compliance guide covers the platform layer in detail.