EU AI Act AI Ad Disclosure Rules for Advertisers (2026)

Quick answer:since 2 August 2026, Article 50(4) of the EU AI Act has required whoever runs an AI system that produces a "deep fake" to disclose, visibly, that the content is artificially generated. The European Commission's Guidelines, published 20 July 2026, settle the question every AI UGC advertiser hoped to dodge: a realistic AI-generated presenter is a "person" under the definition even if they are not modelled on anyone real, and the Guidelines list "AI-manipulated video featuring a realistic synthetic influencer testing out a sponsored real product" as an example that gets no creative exemption. If your AI UGC ad reaches people in the EU, put a human-readable AI label on it. Machine-readable metadata does not discharge the duty. Fines reach 15 million euros or 3% of worldwide turnover.

What actually changed, and when

The AI Act phases in over years. For advertisers, one article matters and three dates:

That last row is the most misreported thing about this regulation. Plenty of coverage presents 2 December 2026 as a general grace period. It is not. The Guidelines describe it as a "targeted grandfathering rule only with regard to the marking and detection obligations under Article 50(2)" for systems already on the market. Nothing in it defers the labelling duty you carry as an advertiser.

Are you a "deployer"?

Article 50 splits duties between providers, who place the AI system on the market, and deployers, who use it under their own authority. The model makers are the providers. You are the deployer, and the Guidelines are specific about the edges:

• The company is the deployer, not the individual. Where a legal person such as an advertising company uses the system, its employees are not separate deployers, and bringing in contractors or freelancers does not shift the duty either.
• Commissioning an agency can take you out of scope.A company that merely commissions an agency to produce an ad, "without taking decisions and exercising control over whether and how the advertising agency uses AI in the production process, is not a deployer". Brief the agency on which AI tool to use and you have taken that decision.
• Being outside the EU is no defence if you target the EU. Third-country deployers are in scope where they foresee dissemination in the Union, "including by directing or authorising distribution within the Union".

Is an AI UGC ad a deep fake? Almost always, yes

Article 3(60) defines a deep fake, and paragraph 113 of the Guidelines breaks that definition into four cumulative criteria: "(i) resemblance of (ii) existing (iii) persons, objects, places, entities or events (iv) that would falsely appear to a person to be authentic or truthful". The two that people assumed would save synthetic-actor advertising, (ii) and (iii), are precisely the two that do not.

On (ii), "existing", the Guidelines say it is "sufficient for simulated persons, objects, places, entities or events to resemble someone or something that exists, can plausibly exist or could have plausibly existed in reality". Read that clause to its end. What falls outside is the physically impossible: content that defies the laws of nature, or "dragons, or elephants driving cars".

On (iii), "persons", the Guidelines are explicit that the word covers realistic human beings "including digital replicas of real persons, realistic AI-generated human avatars or personas, and personal characteristics or expressions, such as image, voice, behaviour, performances etc." An invented, photorealistic creator who exists nowhere but in a model's weights is a "realistic AI-generated human persona". That is the entire category of talking-head AI UGC.

Photorealism alone is not decisive. The Guidelines say it "renders it more likely that such content should be considered a deep fake" without being determinative, and the assessment still turns on capacity to deceive. But they then list "fully AI-generated actors" among the uses of AI that are likely to make content falsely appear authentic.

This is the reverse of the common intuition that an invented presenter is the safe option because they resemble nobody real. Under the Guidelines the invented photorealistic presenter is a central case, not an edge one. Note how that compares with the American rule: New York's synthetic performer disclosure law is the narrower of the two, because it reaches only a performer who is not recognisable as an identifiable real person, while the EU definition covers that case and digital replicas of real people. An invented AI presenter owes a disclosure under both regimes rather than escaping either.

The Commission's own examples, side by side

The Guidelines carry two facing lists, and several entries are literally ecommerce ads. The last row on the right is not from the list itself: it comes from paragraph 116, on manipulations too minor to make content a deep fake. It is included because it is the paragraph advertisers will actually be reasoning about.

Read the right column before you celebrate. Both carve-outs are conditional: the AI backdrop is safe only while the ad does not mislead about the product, and the editing exemption covers only changes with "only a minor impact" on perceived authenticity. A generated hero shot that makes the product look better than the one in the box is expressly on the deep fake side.

The "it's creative work" argument does not survive contact

Article 50(4) has a lighter regime for deep fakes forming part of "evidently artistic, creative, satirical, fictional or analogous work": you still disclose, but in a way that does not hamper enjoyment of the work. Several vendors have floated this as the route for branded content. The Guidelines close it in two moves. First, the categories must be read strictly, and content is excluded from them "if its nature is exclusively informative or commercial and is recognisable as such"; where a piece combines characters, "the informative character should always prevail and require compliance with the standard labelling requirements".

Second, and this is the passage to show anyone who tells you AI UGC is exempt, the Guidelines' list of "deep fakes that do not constitute artistic, creative, satirical, fictional or analogous work" includes:

• "AI-manipulated video in the style of a teleshopping channel involving deep fake simulation of humans advertising a product in an AI-generated scene depicting the functional use of the product by the simulated consumers with the aim of persuading viewers to buy the product."
• "AI-manipulated video featuring a realistic synthetic influencer testing out a sponsored real product, focused solely on displaying the product functionalities."

The second bullet is a definition of the AI UGC ad format. Label it normally.

What a compliant disclosure looks like

Article 50(5) sets the horizontal rules, and the Guidelines put practical shape on them.

• Human-perceivable, not machine-readable.Deployers "cannot rely on the machine-readable marking embedded in the content by the provider under Article 50(2)", because those markings "are not immediately clear and distinguishable" to viewers. The Guidelines point to visible or audible labels that work without "any specific technical tools or performing dedicated actions".
• Clear and distinguishable.A disclosure fails where it "can be easily overlooked or missed by natural persons under normal exposure or interaction conditions". The Guidelines name two failures by example: burying it under menu layers, and putting it in terms of use "that are often not read".
• At first exposure, and possibly again later.The duty runs per output, per viewer. Where it is "reasonably foreseeable that persons may not perceive content from its beginning", a label only at the start is not enough and should be complemented later. For short-form ads people scroll into mid-clip, a persistent on-screen label is the safer build.
• Accessible. Article 50 adds no accessibility rules of its own, but the Guidelines direct you to check whether the EU accessibility directives cover your service and, if so, to make the disclosure conform.

The Commission has also published a Code of Practice on Transparency of AI-Generated Content(10 June 2026), which the Guidelines call "the only Union-wide recognised practical framework" for demonstrating compliance, plus a set of icons deployers may use. Signing is voluntary. Non-signatories "are expected to demonstrate how they have complied" by other adequate means, should run a gap analysis against the Code, and "will likely need" to answer more detailed information requests.

Where the marking duty actually sits, and one thing we can tell you first-hand

Article 50(2) puts the machine-readable marking duty on the provider, not on you. In practice that means the model makers. Google DeepMind's Veopage says "videos made with Veo will be marked with SynthID", and its SynthID page describes that mark as an invisible digital watermark added to the AI-generated image or video segment, "designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression".

Here is the part we can speak to directly, because it is our own code. UGC Vids AI runs a dozen models, remuxes every generated clip on ingest, and re-encodes through ffmpeg on export whenever the file passes through our editor, to libx264 at CRF 23 with -movflags +faststart. Neither path passes -map_metadata, so container-level metadata is not deliberately carried across. We have since measured what that costs. Injecting a C2PA manifest into a test clip at the position the C2PA 2.1 spec mandates and running it through both of our ffmpeg commands removed it every time, including the lossless stream copy that never decodes a frame, while ordinary container tags written with -metadata came through untouched. Pixel-domain watermarks such as SynthID are a different mechanism and that test does not touch them. The measurement, box by box and byte by byte, is written up separately. The compliance point holds either way: a surviving machine-readable mark would not discharge your obligation, and a stripped one would not create it.

One relief for production teams: Article 50(2) does not reach outputs "only used in closed loop environments in industrial and product development workflows (for example for film, animation, games or advertising production), unless they are the final output". Your intermediate generations are not the regulated artefact. The ad you ship is.

Enforcement and exposure

Article 99(4)(g) applies to Article 50 breaches: up to 15,000,000 euros or, for an undertaking, up to 3% of total worldwide annual turnover in the preceding financial year, whichever is higher. SMEs and start-ups get the lower of the two. Supervision sits with national market surveillance authorities, which can act on their own initiative or on a complaint any affected person has the right to lodge.

The AI Act displaces nothing. The Guidelines note that the deep fake test is to be understood "independently and distinctly" from deception under the Unfair Commercial Practices Directive, which applies in parallel. Clearing Article 50 does not clear misleading-advertising law.

A practical checklist

• Decide whether any EU audience is foreseeable for the creative. If yes, assume you are a deployer.
• Sort creative into three buckets: synthetic humans, AI-altered product depiction, cosmetic AI editing. The first two need a label.
• Burn a visible AI label into the file rather than relying on a platform toggle or metadata, and keep it on screen rather than only in the first frame.
• Do not argue creative exemption for a product-demo ad. The Guidelines list that exact format as ineligible.
• Decide deliberately whether to adhere to the Code of Practice or to document equivalent measures of your own.
• Leave pre-2 August 2026 creative alone unless relabelling is cheap. There is no retroactive duty.
• Log which ads used AI and how. If a complaint arrives, the record is your defence.

None of this is legal advice, and the Guidelines call themselves "a first interpretation" the Commission will review. If EU spend is material to your budget, have counsel read Section 6 of the Guidelines against your actual creative.

Sources

• European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689, C(2026) 5054 final, 20 July 2026 (read as the published PDF)
• EU AI Act, Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems
• EU AI Act, Article 99: Penalties
• European Commission, Transparency obligations under Article 50 of the AI Act (Q&A)
• European Commission, Code of Practice on Transparency of AI-generated Content (published 10 June 2026)
• Google DeepMind, SynthID and Veo

Every source above was read directly in August 2026 and every quotation is verbatim. The Commission has said it will revise the Guidelines; the C(2026) 5054 reference above is your check for whether this page has gone stale.

The platforms layer their own rules on top of the statute, and they do not match it. TikTok's AI content disclosure rules put the labelling duty on you, while Meta's AI-generated creative ad policy mostly labels the ad for you. And what makes AI UGC look fake covers the problem a label cannot solve. UGC Vids AI is where the ad gets made in the first place: twelve models on one credit balance, plans from $49/month for 5,000 credits, free for 3 days.