In short
A technical, machine-readable signal embedded in AI-generated content so that its origin can be detected: metadata or an invisible watermark. Article 50(2) of the AI Act requires providers of generative AI systems to apply this marking.
ExampleAn image produced by an image generator carries, in its metadata, a signed statement of its artificial origin. A verification tool can read it before publication.
Why it mattersMarking makes it possible to identify AI-generated content after it has been shared. In some cases, Article 50(4) also requires professional users to add a visible disclosure.
Marking and disclosure: two obligations
Article 50 of the AI Act sets out two separate obligations.
Paragraph 2 applies to providers, meaning the companies behind AI systems that generate audio, images, video or text. They must mark the outputs in a machine-readable format, so that they can be identified as generated or manipulated by AI.
Paragraph 4 applies to deployers, meaning professional users of these systems, such as a newsroom. They must tell the public that a deepfake, or a text published to inform the public on matters of public interest, has been generated or manipulated by AI.
The exceptions
Marking is not required when the AI performs an assistive function for standard editing. Nor is it required when the AI does not substantially alter the input data provided by the deployer or its meaning.
For a text, the visible disclosure is not required if the text has undergone human review or editorial control, and if a natural or legal person holds editorial responsibility for its publication.
The timeline
Article 50 has applied since 2 August 2026. Regulation (EU) 2026/1744 of 8 July 2026 gives providers extra time. Their systems placed on the market before 2 August 2026 must comply with paragraph 2 by 2 December 2026 at the latest. This extension does not cover deployers’ obligations.
The code of practice
The final version of the code of practice on the transparency of AI-generated content was published on 10 June 2026. The European Commission and the European Artificial Intelligence Board have recognised it as an adequate tool for demonstrating compliance. Signing up to the code is voluntary. Article 50 applies in all cases.
For audio, images, video and text contained in a file, signatories apply at least two layers of marking: signed, timestamped metadata and an invisible watermark. Plain text cannot carry metadata. A watermark is enough for it, except for very short texts. Signatories may add a fingerprint or a log of the content produced.
For a published text, the code provides for an icon or an equivalent label. It goes above or at the top of the text, near the headline, or in the notices at the start of the text. The code also allows the disclosure to cover only the part of the text that was generated or manipulated by AI.
Marking in Verbatim
Verbatim records the origin of every line of a text: the journalist, the AI or a source. The files it produces carry machine-readable marking: DOCX, PDF, HTML and content sent to the CMS. This marking lists every passage generated or modified by AI, with its fingerprint. It also gives the share of the text written by AI. A sentence rewritten by the journalist drops out of the marking. Text copied without formatting carries no marking. The record of the origin of every line is ready when the AI Act requires it. In VerbatimOS, the AI Act report covers the whole team.
What a newsroom must do
A newsroom that uses a provider’s tool is a deployer. Technical marking is the provider’s responsibility. The newsroom must tell the public about deepfakes it produces with AI. It must also disclose AI-generated or AI-manipulated texts on matters of public interest, unless the editorial exception applies.
A newsroom that develops its own tool and puts it into service under its own name becomes a provider. It must then mark the content that tool produces.
Questions
Does a newsroom have to mark articles written with AI?
Technical marking is the responsibility of the tool’s provider. The newsroom must disclose a text on matters of public interest generated or manipulated by AI. This disclosure is not required after human review or editorial control, under the editorial responsibility of a natural or legal person.
What is the difference between marking and labelling?
Marking is a technical signal that a machine can read. Labelling is a notice the public can see, such as an icon or a sentence at the top of the text.
Since when has marking been mandatory?
Since 2 August 2026 for systems placed on the market from that date. Systems already on the market have until 2 December 2026.
What are the penalties for non-compliance?
Up to €15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs and small mid-cap companies, the lower amount applies.
Sources
- Regulation (EU) 2024/1689 on artificial intelligence (AI Act), Official Journal L of 12 July 2024 (in French)
- Regulation (EU) 2026/1744 of 8 July 2026, known as the digital omnibus on AI (in French)
- European Commission, guidelines on the transparency obligations of providers and deployers of AI systems
- European Commission, code of practice on the transparency of AI-generated content