AI content labeling under Article 50 of the AI Act: rules from August 2026

From 2 August 2026, anyone publishing AI-generated content for audiences in the EU is responsible for labeling it visibly — regardless of what the tool vendor does. That is the effect of Article 50 of the AI Act (Regulation EU 2024/1689), the transparency rules for limited-risk AI systems. The duty covers the tools marketing teams use daily: text, voice, image and video generators. Below we break down which materials need a label, what the label must look like, and where relying on a platform's disclosure will not protect you.

The obligation applies to anyone creating or distributing AI content aimed at EU audiences — the company's registered seat is irrelevant. That includes agencies and in-house teams inside the EU, but also US or UK brands whose campaigns reach European users. The regulation calls this role the deployer: you use an AI tool and publish the output, so you own the disclosure. Neither the platform nor the model provider takes that duty off you.

Everyday marketing tools — text, voice and image generators — fall into the limited-risk tier, and Article 50 defines the labeling duties precisely for that tier. The AI Act sorts systems into four levels, with requirements growing alongside risk. Knowing the full split tells you where your work starts and stops being regulated:

You own the visible label — the tool provider only owns the technical watermark. This is the most common misunderstanding around Article 50: because the tool „marks something itself”, the publisher assumes the duty is covered. The regulation, however, separates the two roles and imposes two independent, parallel obligations. Meeting one does not satisfy the other.

The AI system provider (for example, a voice or image generator vendor) must embed metadata and a watermark in the file so machines can detect the content is synthetic. That marking is invisible to your audience — it is read by platform algorithms and detection tools. From the viewer's perspective, the file looks like any ordinary photo or recording. That is exactly why the provider's watermark does not satisfy the duty to inform the audience.

Your duty is a visible, human-readable disclosure: a label on the image, an overlay on the video, a spoken notice in audio, a note next to the text. It applies independently of what the vendor did technically. If the voice generator embeds a watermark but you skip the „AI-generated voice” notice, the obligation remains unmet. For multi-country campaigns, localize the label into each market's language — a Polish audience should read the disclosure in Polish, a German one in German.

Two categories require a label: realistic synthetic media (images, audio, video that could pass for genuine) and texts informing the public on matters such as politics, the economy or health, when published without human editorial review. Outside those categories, Article 50 requires nothing — a product description polished with a text generator or a comic-style illustration stays unlabeled. The dividing line is the risk of misleading the audience about authenticity.

Label every asset a viewer could mistake for a real recording or photograph. There is no exemption for „well-made” material — the more realistic it is, the more clearly the duty applies. In day-to-day marketing work this covers:

You do not need to label minor edits that leave the character of the content intact, nor material that is obviously non-realistic. Doubtful cases are settled by a simple test: show the asset to someone outside the project and ask whether they believe it is a real photo or recording. Out of scope are, among others:

You can use the European Commission's official icons or your own brand-consistent label — both satisfy the requirement as long as the marking is legible and visible. The Commission prepared three icons: AI GENERATED (fully AI-made content), AI MODIFIED (AI-processed content) and a base AI mark. The choice depends on how deeply AI intervened in the asset. What matters legally are the legibility parameters, not the icon design itself.

Design the company label once, add it to your asset library, and it works in every campaign. It must, however, meet three conditions at the same time — otherwise it fails as a disclosure under Article 50:

Placement depends on the format, but the principle is constant: the audience must see or hear the disclosure without an extra click. Hiding the information in metadata, behind a „more” link or inside a menu does not meet the visibility requirement. The standard looks like this:

On Meta the line runs between organic posts and ads: for the former the platform's label is enough, for the latter it is not. It is a practical trap many teams across the EU fall into, because the mechanism looks identical while the legal outcome is the opposite. Both scenarios below.

For organic posts, an automatic disclosure such as „Made with AI”, shown directly under the post, satisfies the visibility requirement. The platform acts as a detection and labeling layer, and the viewer sees the notice without any action. Still, verify the label actually appeared — automatic detection is unreliable with mixed content. If the platform missed it, the duty falls back on you.

In ad campaigns, Meta tucks the AI disclosure behind the three-dot menu — the viewer must click to see it, so it is not a „visible label” within the meaning of Article 50. The consequence: embed the AI label directly in the creative, on the graphic or video, in a visible spot. This applies to every campaign using AI-generated material, in every EU market, regardless of budget or format. We check this as part of our ad account audits for clients.

The rollout comes down to five steps: tool inventory, a labeling policy, a company label, updated briefs and an ad audit. A marketing team needs a day or two for the whole thing — and it removes the risk from every future publication. Below is the order we recommend to clients.

Start with a list of AI tools used across the team and mark which of their outputs reach publication. Then write a one-page labeling policy: which content needs a label, what the company label looks like, and where it goes per format. Design the label once — meeting the three conditions above — and add it to the asset library. For multi-market brands, prepare localized label variants per language alongside the visual asset.

Add a fixed checklist to your content approval flow — labeling should be routine, not an exception. Before every publication, answer in order:

Teams that roll out labeling before 2 August avoid fixing live campaigns under pressure and build audience trust before the market forces it. Transparent AI disclosure is also becoming a quality signal for the models themselves — AI search and recommendation systems favor sources that clearly declare content provenance. That is the same direction we take clients in with our SEO and LLM visibility services. If you want to run the rollout with us, get in touch — we will draft a labeling policy for your channels and markets.