Since 2 August 2026, Article 50 of the EU AI Act requires anyone who uses AI to generate or manipulate a deepfake (realistic image, audio or video that resembles existing people, objects, places or events and could falsely appear authentic) to disclose that it is AI-generated, clearly and at the latest when the viewer first sees it. Not every AI-assisted ad needs a label. Realistic ones almost always do.
Not legal advice. For a specific campaign, ask counsel in your market.
What EU AI Act Article 50 says
Article 50 sets transparency duties for providers (who build the AI system) and deployers (who use it), applying from 2 August 2026 (AI Act Explorer). Two paragraphs matter for ads (Article 50 text):
| Paragraph | Who | Duty |
|---|---|---|
| 50(2) | Providers | Output must be "marked in a machine-readable format and detectable as artificially generated or manipulated" |
| 50(4) | Deployers | Disclose that deepfake image, audio or video is AI-generated or manipulated |
| 50(5) | Both | Disclosure must be "clear and distinguishable," given "at the latest at the time of the first interaction or exposure" |
Provider vs deployer in an ad
The tool maker is the provider and must machine-mark the output. Under the AI Omnibus (adopted by the Council on 29 June 2026, per iubenda), generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet this marking rule. Newer systems get no grace period, and deployer duties still apply from 2 August.
Whoever uses the tool to make the content is the deployer, and owns the visible label. Per Mishcon de Reya, the Commission's final guidelines say a brand that merely commissions an agency, without deciding how it uses AI, is not the deployer; the agency is. Put the duty in the contract either way.
Is your ad a "deepfake"?
The Act's definition: AI content that "resembles existing persons, objects, places, entities or events" and would falsely appear authentic. The final guidelines (20 July 2026) read it widely, per Faegre Drinker and Mishcon:
- It is enough that something "could plausibly have existed." Intent is irrelevant.
- More photorealism makes a deepfake finding more likely.
- Fully AI-generated actors and digital replicas are likely deepfakes.
- Background scenes and special effects are unlikely to be.
- AI marketing content that makes products appear different from reality counts.
Our reading:
| Scenario | Likely deepfake? |
|---|---|
| Photoreal AI person presenting the product | Yes |
| AI shot making the product look bigger or glossier than it is | Yes |
| Realistic "fake out-of-home" stunt on a recognizable street | Likely |
| Real product shoot with an AI-extended background | Unlikely |
| Clearly impossible stylized world (chrome bottle morphing in a void) | Unlikely |
For content that is part of an "evidently artistic, creative, satirical, fictional or analogous work or programme," the duty shrinks to disclosing that such content exists in a way that "does not hamper the display or enjoyment of the work." It never disappears, and per Faegre Drinker the guidelines let most advertising examples fall outside this lighter regime.
What Article 50 does not require
- A label on every AI-assisted ad. Grading, cleanup and clearly non-realistic content are not deepfakes by default.
- Fixed wording. The voluntary Code of Practice, final since 10 June 2026, points to "the publicly available EU icon or equivalent labels" (Jones Day).
- Retroactive labels on content published before 2 August 2026 (Mishcon).
- A human-review exemption for video. That exemption covers only AI text on public-interest matters.
Fines for breaching Article 50 reach €15 million or 3% of worldwide turnover, whichever is higher; for SMEs, whichever is lower (Article 99).
Platform AI labels are a separate layer
- Meta adds "AI info" to ads made or significantly edited with its AI tools, and detects third-party AI (Meta).
- YouTube requires disclosure of realistic synthetic content and labels photoreal content on the player (YouTube Help).
- TikTok requires labels on realistic AI images, audio or video (TikTok).
A platform toggle does not satisfy the Act, and vice versa. Do both.
AI generated ads label: a checklist
- Log AI use per asset: which shots, which model, generated vs real.
- Run the deepfake test on each asset.
- If yes, label at first exposure: an "AI-generated" super or the EU icon in the opening frame, plus the caption. Audio-only gets a spoken disclaimer.
- Keep provenance metadata; check what your encoder strips.
- Tick each platform's AI disclosure on upload.
- Write the duty into contracts and archive the log.
Budgeting? See AI video ad costs and AI vs traditional production. Honest labeling also helps the work, as we argue in why most AI commercials look bad.
