The AI label is entering the publishing workflow

The small label is about to carry a lot of work. For years, platforms and publishers have talked about marking AI-generated content as if the hard part were choosing the right badge. Europe’s latest move makes the question more operational: who adds the label, what exactly is being labelled, where the signal sits, and whether a machine can still detect it after the content has been copied, compressed, cropped or reposted.

On 10 June, the European Commission published a final Code of Practice on marking and labelling AI-generated content. The code is voluntary, but the legal duties it is meant to support are not. Article 50 of the EU AI Act will bring transparency obligations into application from 2 August 2026. The Commission says the code is designed to help providers and deployers of generative AI systems meet those obligations in a consistent and practical way.

That distinction matters. A voluntary code sounds soft until it meets a mandatory rule. Providers of systems that generate synthetic audio, images, video or text are expected to mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated, as far as technically feasible. Deployers have a different problem: in key cases, they must disclose that people are seeing a deepfake or certain AI-generated or AI-manipulated text published to inform the public on matters of public interest.

The visible label is only the part a reader notices. Behind it sits a product workflow. A model provider has to decide how a generated image, voice clip, video or block of text can carry a durable signal. A platform has to decide how it will preserve, read or display that signal when content moves through its systems. A publisher, campaign group, creator or business using generative AI has to decide when a human review, editorial responsibility or a public-facing disclosure changes what the audience needs to see.

The EU’s optional icon set shows how specific the problem has become. The Commission has produced icons for labelling certain AI-generated or AI-manipulated content, with versions for fully AI-generated material and partially AI-modified material. Its own summary notes that user testing improved when the basic icon was accompanied by a text label, such as “modified”. That is a useful warning against badge minimalism. A tiny symbol may satisfy a design team, but it may not tell an ordinary reader enough.

The rules are also narrower than a panic headline would suggest. The Commission’s icon guidance says the AI Act disclosure requirement for public-interest text applies where the text is AI-generated or AI-manipulated, published to inform the public, did not undergo human review or editorial control, and no natural or legal person assumed editorial responsibility. That does not make every spell-checked paragraph or assisted draft a public label case. It does make the absence of a real editorial workflow harder to hide behind.

For chatbot and product teams, the same transparency logic reaches the interface. Article 50 also requires people to be informed when they are interacting directly with an AI system, unless that is obvious in the circumstances. The practical version is not a legal paragraph buried in terms of service. It is the first screen, the chat header, the support widget, the voice assistant prompt and the handover to a human when the system has reached its limits.

The more difficult test will come after launch. Labels can be stripped by screenshots, watermarks can be cropped, metadata can be lost and synthetic media can move faster than moderation queues. The code leans towards interoperability and detectable marking, but it does not make the technical problem vanish. It also does not replace future Commission guidelines on Article 50. Anyone selling certainty here is moving faster than the rulebook.

Still, the direction is clear enough for planning. AI disclosure is becoming less like a footer and more like a supply chain. The model creates a signal, the product preserves it, the platform displays it, the publisher explains it, and the audience gets a chance to understand what kind of content is in front of them. The best implementations will probably feel boring. They will not present labelling as a dramatic confession. They will make it part of the ordinary route from generation to publication.

That is the useful shift. The AI label is not just a badge stuck onto the end of a story, advert, video or chatbot. It is a record of decisions made earlier: whether the material was synthetic, whether it was changed, whether it was reviewed, whether someone took responsibility, and whether the person seeing it can tell the difference without becoming a forensic analyst.