From 2 August 2026 the transparency rules in Article 50 of the EU AI Act apply, and most of the coverage aimed at marketers gets the scope wrong. If you are a brand running AI-assisted content you are a deployer, not a provider, and the deployer obligation is narrow: disclose content that qualifies as a deep fake, and disclose AI-generated text published to inform the public on matters of public interest. A stylised product animation is neither. A synthetic presenter who looks like a real person is the first one.
This guide covers which role you actually hold, the three-part test that decides whether an asset needs a label, how the label should look under the Commission's Code of Practice, what belongs in the creator brief, and the cases where this is a ten-minute policy rather than a project.
This is not legal advice. Every rule below is cited to its article or source so your counsel can check it against what you are actually publishing.
Provider or deployer: the distinction that sets your scope
Article 50 splits obligations by role, and brands consistently read the wrong half.
Providers develop an AI system and place it on the EU market. Under Article 50(2) they must ensure the system's synthetic audio, image, video and text outputs are marked in a machine-readable format and detectable as artificially generated. That is your tool vendor's obligation, not yours. Generative systems already on the market before 2 August 2026 were given until 2 December 2026 to meet that marking requirement; systems placed on the market from 2 August 2026 onward get no grace period at all.
Deployers use an AI system under their own authority. That is you, every time your team or your agency generates a video, an image or a voice-over for a campaign. Your obligations live in Article 50(4), and there are only two of them.
The practical consequence: ask your vendor whether their outputs carry machine-readable provenance, because it is their legal duty and it makes your own disclosure easier to evidence. But do not build a compliance programme around a paragraph that does not address you.
The test that decides everything: is it a deep fake
Article 50(4) requires deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake to disclose that the content has been artificially generated or manipulated. Everything therefore turns on the definition in Article 3(60), which sets three cumulative criteria.
Resemblance. The content closely resembles a person, object, place, entity or event.
Existing. What it resembles exists, plausibly exists, or could plausibly have existed in reality.
False appearance of authenticity. It would falsely appear to a viewer to be authentic or truthful.
All three must be met, and applied to what a brand actually ships that produces clear answers. A synthetic presenter with a realistic human face and a cloned voice demonstrating your product: yes, all three. A photoreal AI shot of your product on a kitchen counter that never existed: yes. An obviously illustrated or animated explainer nobody would mistake for footage: no, the third criterion fails. A real creator video that went through AI colour correction or upscaling: no, and Article 50(2) itself carves out assistive editing that does not substantially alter the input data or its semantics.
The obligation is not limited to editorial contexts. Commission guidance is explicit that it reaches commercial promotion — websites, catalogues, billboards, TV and social media — so "it is only an ad" is not a defence.
The text obligation that almost never applies to brands
The second limb of Article 50(4) covers deployers of systems generating or manipulating text published with the purpose of informing the public on matters of public interest. That is a narrow category, and a product description, a caption or an email subject line is not in it. There is also an explicit exemption where the content underwent human review or editorial control and a natural or legal person holds editorial responsibility for the publication.
Read plainly: unless you are publishing AI-written journalism or commentary on matters of public interest, this limb does not reach your marketing copy. Several vendor blogs claim that all AI-written content must be labelled from August 2026. That is not what the article says.
Your chatbot is a separate rule
Article 50(1) requires that AI systems intended to interact directly with people be designed so the person is informed they are dealing with an AI system, unless it is obvious from the context. The drafting puts that duty on providers, but it shapes what you may reasonably deploy: a branded assistant that lets a customer believe they reached a human is the exact outcome the article exists to prevent. Say it once, in the first message, in plain language. TikJoy's WhatsApp concierge identifies itself at the start of a conversation for this reason, and it costs nothing in conversion when the assistant is actually useful.
What a compliant label looks like
The Commission published the Code of Practice on Transparency of AI-generated Content on 10 June 2026, confirmed it adequate together with the AI Board in July 2026, and published the final version of its Article 50 Guidelines on 20 July 2026. The Code is voluntary: signing gives you a recognised route to demonstrating compliance, while non-signatories have to evidence it another way and can expect closer scrutiny from market surveillance authorities.
Under the Code a deployer applies a clear visual label at the point of first exposure to the content. The Guidelines indicate that the Commission's standardised icon satisfies this requirement, and permit alternatives meeting the Code's specifications. Where a visual label is impractical — audio-only content, for instance — an equivalent spoken or written disclaimer is required instead.
Two things brands get wrong here. First, "at the point of first exposure" means the viewer sees it when the content starts, not in a caption they have to expand or a legal page they never open. Second, a platform's own AI toggle is a platform rule, not proof of Article 50 compliance. Use both: the toggles are covered in our guide to labelling AI-generated UGC on TikTok and Meta, and they stay necessary for distribution reasons regardless of what EU law requires.
What changes in the creator brief
None of this is manageable if you cannot tell which assets are synthetic, and with commissioned content you did not press the button.
Ask which tool produced the asset and keep the answer. Provenance metadata from a compliant provider is the cheapest evidence you will ever collect.
Ask specifically about faces and voices. The question is not "did you use AI" but "is any human face or voice in this synthetic or cloned". That is what the deep-fake test actually turns on.
Set the label as a deliverable, not a preference, with the wording and placement you want, so it is applied before you ever see the file. Our UGC brief template has the fields, and the rights and licensing terms should say who carries the disclosure duty when you reuse the asset in paid media.
When this is the wrong thing to worry about
If your programme runs on real creators filming real products, Article 50 is a short written policy and a line in a brief, not a workstream. Spending a quarter building an AI-disclosure framework while your actual exposure sits elsewhere is a real misallocation.
For most brands the higher-probability risks are the ordinary ones: misleading-advertising and unfair-commercial-practice rules that apply to the claim whether or not a machine drew the picture, and platform policies that cut your distribution long before a regulator writes to you. In Italy there is also Law 132/2025, in force since 10 October 2025, whose new Article 612-quater of the criminal code punishes the harmful, non-consensual distribution of AI-falsified images, video or audio of a person with one to five years' imprisonment. That provision is about real people depicted without consent, not about your product photography, and conflating the two produces bad decisions in both directions.
Penalties under the AI Act itself for breaching Article 50 reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher. That number is what makes the topic loud. The scope above is what makes it manageable.
The practical position
Write down, once, which of your content types can contain a synthetic human or a photoreal scene that never happened. Label those at first exposure, in the platform toggle and visually. Ask vendors for machine-readable provenance and creators for a straight answer on faces and voices. Leave your product copy alone.
That is close to the whole obligation for a brand, and it is why disclosure belongs in the brief rather than bolted on at review. For the vertical-by-vertical view of which rules bite your category, see your industry's rules.