AI Act, Article 50: marking of AI visuals is now in force. What your brand needs to do now

Are you in scope?

Article 50 of the European AI regulation deals with the transparency of content produced or edited by AI. The criterion is not the size of your company, it is your usage: as soon as AI-generated images, videos or audio reach the public, the obligation exists. A few very common situations are enough to fall within scope:

• Packshots and e-commerce visuals generated by AI for your product pages;
• Social media content produced with an image-generation tool;
• Photos substantially edited by AI: background replaced, product cut out and restaged;
• Multi-country campaigns automatically adapted by generative models.

Marking AI visuals is no longer a best practice: it is an obligation.

What Article 50 requires

The text asks for two distinct things. First, telling people when they are interacting with an AI, a conversational assistant for instance. Second, and this is what concerns your visuals: any content generated or manipulated by AI must carry a machine-readable mark, embedded in the file, so that a system can automatically identify it as artificial. A watermark visible to the eye is not enough; neither is a logo placed on the image.

Responsibility runs along the chain: the AI system provider marks the outputs, and the brand that publishes must make sure its content actually carries the mark. Breaches expose you to fines of up to 15 million euros or 3% of annual worldwide turnover.

The dates that matter

The timeline is easy to remember, with one nuance introduced this summer by the EU’s Digital Omnibus package on AI:

• 1 August 2024: the AI Act enters into force. Obligations have been phased in since then.
• 2 August 2026: Article 50’s transparency obligations apply. User information and content labelling are enforceable, and any generative AI system placed on the market from this date must mark its outputs in a machine-readable format.
• 2 December 2026: end of the grace period granted by the Omnibus (in force since 27 July 2026) to systems already on the market before 2 August. After that date, no exceptions remain.

In other words: if your generation tool is recent, the obligation already applies; if it was on the market before 2 August, it has a few months left to catch up. In both cases, the question to ask your provider is the same: are your outputs marked, and how?

C2PA and SynthID: what a compliant visual looks like

Two technologies dominate. The C2PA standard attaches a provenance manifest to the file: who produced the content, with which tool, and what was modified. Platforms and verification tools can read it. SynthID, developed by Google DeepMind, embeds an invisible watermark directly in the pixels of content produced by Google’s models: the eye sees nothing, a machine detects it.

At PixAi, marking is not an option to switch on: PixAi Frame builds C2PA marking into its production pipeline, complemented by SynthID whenever generation relies on Google’s models. Compliance shows in the files, not in the sales deck. What remains is putting it in place on your side, and it is simpler than it looks.

Getting compliant: where to start

• Map it out: identify where your teams generate or edit visuals with AI, including through agencies and freelancers.
• Check the actual marking: request a delivered file and inspect what it carries. An “AI-generated” badge placed on the image is not machine-readable marking.
• Keep records: archive what you produce, with which tool and which version. That is the basis of sound governance, beyond compliance.
• Prefer tools that mark by default: a manual step always ends up skipped on a rush day.

Article 50 is not a wall. If your production chain marks your visuals on its own, compliance becomes an observation rather than a project. And if you want to see concretely what it looks like on your own products, we can generate a few together in thirty minutes.

This article reflects our operational reading of the text; it does not replace legal advice.