If you run ads, someone will ask a new question: did AI make this? In the EU it is no longer a question about taste: it has dates and fines attached, and it reaches companies with no office there.
What changed, and when
The transparency rules in the EU AI Act — Article 50 — apply from 2 August 2026 1. The rewrite did not touch them: the Digital Omnibus on AI, in force since 27 July 2026, pushed high-risk obligations for stand-alone systems to 2 December 2027 and for product-embedded systems to 2 August 2028, and left Article 50 applying from 2 August 2026 2.
One grace period exists, and not for advertisers: systems on the market before 2 August 2026 have until 2 December 2026, for the provider-side marking duty only 1. The Commission's FAQ puts the penalty at up to 15 million euros or 3% of total worldwide turnover for the preceding financial year 1. Distance does not help: the Act covers providers and deployers in a third country where the output is used in the Union 3.
Who marks, and who labels
Article 50 splits the work in two, and it matters which side you are on. The provider is the company whose system made the output; the deployer is whoever used it. Buy a subscription to an image model: you are the deployer. Providers must ensure synthetic output is, in the Act's words:
“marked in a machine-readable format and detectable as artificially generated or manipulated”
That falls away where the system only assists standard editing, or does not substantially alter the input or its semantics. Marking is machine-side: metadata a customer never sees. The visible part is the deployer's: they must disclose that deep fake content has been artificially generated or manipulated, lightened for evidently artistic, creative, satirical or fictional work 4. Deployers must also disclose AI text published to inform the public on matters of public interest, unless a person holding editorial responsibility reviewed it. Disclosure is owed at first exposure 4.
| The duty | Falls on | The carve-out |
|---|---|---|
| Machine-readable marking of synthetic output | Provider of the system | Assistive editing that does not substantially alter the input |
| Visible disclosure of a deep fake | Deployer — whoever made the ad | Lightened for evidently artistic, creative or satirical work |
| Disclosure of AI text on matters of public interest | Deployer | Review by a person holding editorial responsibility |
| At first interaction or exposure | Both | None |
A voluntary Code of Practice on transparency of AI-generated content followed on 10 June 2026: a provider section on marking, a deployer section on labelling, a standard EU icon set 5. Around 190 organisations had signed by late July 2026; the rest must show authorities their own measures are adequate 5.
What probably needs a visible label
Honest advice thins here. Article 50(4) attaches the visible duty to deep fakes 4. The ends are clear: a drawn illustration is a weak candidate, a photoreal person who never existed saying words nobody said is a strong one. Most work sits in between.
- Likely in scope: a photoreal presenter who never existed, a cloned voice that sounds like a real person, a scene built to look like footage of a real place.
- Likely outside: retouching, colour grading, background clean-up — assistive editing 4.
- Unclear: an invented stock-style model, an AI picture of a real product, a headline a model drafted and a person rewrote.
The platforms, and three other rulebooks
Platforms write their own terms, and they disagree. TikTok's advertising policy, last updated April 2026, requires significantly edited media and AI-generated content to carry its AIGC label or the advertiser's own disclaimer, caption, watermark or sticker; it treats AI voice-cloning as a substantial modification, and says an undisclosed AI ad will be rejected or restricted 6. That is a contract, not a statute; one platform tells you nothing about the next.
New York aims at ad production. S.8420-A amends section 396-b of the General Business Law, was signed on 11 December 2025 and takes effect 180 days after becoming law 7. Whoever produces a commercial advertisement must conspicuously disclose a synthetic performer — an AI-made asset resembling a human performer who is not a recognisable real person — where they have actual knowledge of it. The penalty is $1,000 for a first violation, $5,000 after 7.
California's AI Transparency Act became operative on 2 August 2026, covering generative-AI services with more than one million monthly users in California, at $5,000 per violation with each day counted separately 8. What reaches an advertiser is downstream: providers must require licensees by contract to keep the embedded latent disclosure working, and revoke the licence within 96 hours if it is disabled. That is a law firm's summary, not the statute 8.
The UK is loosest. The ASA and CAP wrote in May 2025 that there is no across-the-board duty to say an ad used AI, that the test is whether the audience would be misled without it, and that a disclosure rarely repairs a fundamentally misleading message 9.
The US Federal Trade Commission has narrowed its aim without going quiet: on 22 December 2025 it set aside its order against the AI writing tool Rytr, while saying it will still hold accountable those who use AI to break the law or mislead consumers about what their generative AI can do 10. Reviews are the other surface. The Consumer Reviews and Testimonials Rule took effect on 21 October 2024 11; liability can attach where a business knew or should have known, and the FTC's own questions and answers say advertising agencies and PR firms are not immune, while the rule carries no blanket ban on AI avatars 11. The invented experience behind one is the problem.
Five steps for a small team
None of this needs a compliance department, just a habit: record-keeping.
- Write down what AI made. Three columns on the creative sheet you already keep: tool and version, date, whether a person changed it. Every rule above turns on a fact nobody remembers later.
- Label realistic synthetic people, voices and events on the asset, not on a policy page: the EU text asks for it at first exposure, not on click 4.
- Keep a human in the loop and keep the evidence: review by a named person holding editorial responsibility is what switches off the text duty in Article 50(4) 4.
- Substantiate the AI claims in your own marketing: if you call a tool accurate, or a replacement for a person, file the evidence before the sentence goes live 10.
- Ask vendors what they mark. Providers carry the marking duty 4, and some must revoke a licence from anyone who strips their embedded disclosures 8. Check your exports keep the metadata.
What stays unclear is worth naming: where assistive editing ends, how much a person must change before an asset stops counting as synthetic, whether the platforms converge on one label. If your ads run in more than one market, ask a lawyer who knows them.
Where the Troy Score fits
The free Troy Score is an independent read of a marketing account, graded on measurement, creative, targeting and pacing. It is not a compliance audit and will never tell you that you are compliant. It does look at creative: how many concepts are live, and where they came from. A team that cannot say where an asset came from has a record-keeping problem before a disclosure one.
Sources
Every claim above points at one of these. Dates are the source’s own.