Article 50 asks for *perceptible* marking, not just machine-readable

Think of a 'contains nuts' sticker. It has to be on the front of the packet where you can see it, not printed on the inside in tiny letters. That is roughly what Article 50(2) of the AI Act asks for when AI-generated content could mislead someone.

The rule is often described as requiring machine-readable marking, which is true but incomplete. The full text says the marking must be in a machine-readable format and in a manner that is perceptible and understandable to humans ^[1]. The second part matters just as much as the first.

What must be marked

Providers of AI systems that generate or manipulate image, audio or video content must mark that content when it resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful ^[1].

The technical standard on this, still being finalised, will spell out how the marking works. Until then, the law itself is clear: machine-readable is not enough on its own when the content could be mistaken for real ^[1].

For text, image, audio and video that do not fall under that threshold, the machine-readable marking suffices ^[1].

The deployer's duty

Deployers have a separate duty under Article 50(4). When they use AI to create or manipulate image, audio or video that amounts to a deepfake, they must disclose that it is artificially generated or manipulated. If it is part of an evidently artistic, creative, satirical or fictional work, and a disclosure would spoil the work, they can disclose the existence of such content in an appropriate manner instead ^[2].

For AI-generated text published to inform the public on matters of public interest, disclosure is required unless a person has reviewed it and a natural or legal person holds editorial responsibility ^[2]. That means a news publisher can use AI to draft an article, but if an editor checks it and the publisher stands behind it, the text does not need a disclosure.

When it applies and who it covers

Article 50 has applied since 2 August 2026 ^[2][3]. It covers providers placing general-purpose AI models with systemic risk on the EU market, and deployers using any AI system that generates or manipulates content ^[2].

A breach can lead to fines up to €15 million or 3 per cent of global annual turnover for the preceding financial year, whichever is higher. For small and medium-sized enterprises, including start-ups, the lower of the two applies ^[3].

What this means for you

If your AI generates realistic images, audio or video, check whether your marking is visible to people, not just encoded for machines. Metadata alone will not meet the rule when the content could be mistaken for reality.

If you deploy AI that creates deepfakes or generates text for public information, decide now who reviews it and who takes responsibility. The disclosure duty is yours, and it applies whether the content is on your own site or shared elsewhere.

For media organisations, the editorial responsibility exemption is useful but not automatic. Someone must actually review the text and the organisation must stand behind it.

What to watch

The European Commission is working on implementing acts that will set out the technical details of the machine-readable marking ^[1]. Those acts will clarify which formats and methods are acceptable, but they will not remove the human-perceptible requirement where it applies.

For now, the law is what you have to work with, and it already says that marking realistic AI content is a job for humans as well as machines.