What the AI Act watermarking deadline means for providers and deployers
The AI Act watermarking deadline on 2 December 2026 is the date by which providers of generative AI systems already placed on the EU market before 2 August 2026 must start marking synthetic audio, image, video and text output in a machine-readable format so it can be detected as artificially generated or manipulated. The duty is set out in Article 50(2) of Regulation (EU) 2024/1689, and the Digital Omnibus (Regulation (EU) 2026/1744) extended the time to meet it for systems launched before 2 August 2026. Content generated and published before 2 August 2026 does not need retroactive labelling.
From 2 August 2026, new generative AI systems must already meet the marking duty when they are placed on the market. Deployers have separate duties under Article 50(4): they must disclose deepfakes and disclose AI-generated or manipulated text published to inform the public on matters of public interest. The Digital Omnibus did not change those deployer duties or their start date.
Who is in scope for the 2 December 2026 marking duty
The 2 December 2026 marking duty applies to providers of generative AI systems that were already on the EU market before 2 August 2026. Providers are the companies that develop or significantly modify a system and place it on the market or put it into service. If your organisation builds or customises a generative AI model and offers it to users in the EU, you are a provider and must add machine-readable markers to its outputs by the deadline.
Deployers are the organisations that use an AI system under their own authority. Deployers do not have the marking duty under Article 50(2), but they do have disclosure duties under Article 50(4) for deepfakes and certain AI-generated text. If you are a deployer, check whether the provider of the system you use has implemented marking, because your disclosure obligations may depend on being able to detect or label the content.
What the Digital Omnibus changed and what it did not
The Digital Omnibus added a four-month transition period for generative AI systems already on the market before 2 August 2026, giving them until 2 December 2026 to meet the Article 50(2) marking duty. It did not change the start date for new systems, nor did it alter the deployer disclosure duties in Article 50(4). It also introduced a prohibition on AI nudification systems from 2 December 2026, but this is a separate obligation and does not affect the marking deadline.
The Omnibus did not move the Article 50(1), (3) or (4) duties. Article 50(1) requires providers to tell people when they are interacting with an AI system; this has applied since 2 August 2026. Article 50(3) requires deployers of emotion recognition or biometric categorisation systems to inform the people exposed, and Article 50(4) requires deployers to disclose deepfakes and certain AI-generated text, both from 2 August 2026.
What machine-readable marking means in practice
Article 50(2) requires that synthetic audio, image, video and text output is marked in a machine-readable format so it can be detected as artificially generated or manipulated. The marking must be effective, interoperable, robust and reliable as far as technically feasible. The Commission’s Guidelines on Article 50, published on 20 July 2026, clarify that a multi-layered approach is recommended: metadata (such as a signed C2PA manifest) combined with an invisible watermark embedded in the content itself.
Metadata can be stripped when content is screenshotted or republished by platforms that remove it, so an invisible watermark provides a fallback. The Code of Practice on Transparency of AI-Generated Content, finalised on 10 June 2026, asks for metadata that is digitally signed and time-stamped in a secure, tamper-evident way. Signatories can rely on the Code to demonstrate compliance with Article 50(2), (4) and (5), but signing is voluntary.
How to meet the marking duty by 2 December 2026
If you are a provider with a generative AI system already on the market before 2 August 2026, you need to implement machine-readable marking before 2 December 2026. Start by auditing your outputs: identify which content types (audio, image, video, text) your system produces and whether they are synthetic or manipulated. Then choose a marking solution that combines metadata and invisible watermarks, ensuring it is interoperable with common detection tools.
Next, integrate the marking into your pipeline so that every output carries the required markers. Test the detectability of the marks across different platforms and use cases, and document your approach. If you use a third-party solution, verify that it meets the technical requirements of Article 50(2). For deployers, ensure that your workflows can detect and label AI-generated content where required by Article 50(4), and that your disclosure is clear, distinguishable and provided at the latest at first interaction or exposure.
Exemptions and practical considerations for creators and studios
There is a narrow exemption from the marking duty for outputs used in business-to-business or industrial contexts. Artistic, creative, satirical or fictional works only need disclosure in a way that does not hamper the display or enjoyment of the work. AI-generated text under human review with editorial responsibility is exempt from the public-interest text disclosure in Article 50(4).
For studios and agencies, this means that internal or B2B use of AI-generated content may not require marking, but public-facing content likely does. If you publish AI-generated images or video for creative projects, ensure that the disclosure does not disrupt the viewer’s experience while still meeting the legal requirement. For text, if your editorial team reviews and takes responsibility for AI-generated articles, you do not need to label them as AI-generated for public-interest disclosures.
What happens if you miss the deadline
Breaching the Article 50 obligations can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher (lower ceilings apply to SMEs). The risk is not just financial: failing to mark content can erode trust with users, partners and regulators. Enforcement actions are likely to focus on high-impact or deceptive uses of AI, so prioritise compliance for content that could mislead the public or influence decisions.
If you are behind schedule, focus first on the most widely used outputs and the highest-risk use cases. Use the Code of Practice as a framework if you need a structured approach, and consider engaging with industry groups to share best practices. The Commission’s Guidelines and the voluntary EU icons for labelling AI content can also help standardise your disclosures.
Frequently asked questions
Does the Digital Omnibus move the watermarking deadline to 2 December 2026 for all AI systems?
No. The Digital Omnibus only extends the Article 50(2) marking duty deadline to 2 December 2026 for generative AI systems already on the market before 2 August 2026. New systems placed on the market from 2 August 2026 must meet the marking duty immediately.
Do I need to comply with Article 50 if my company is based outside the EU?
Yes, if your generative AI system is used by people in the EU or your outputs are published there. The EU AI Act applies to providers and deployers regardless of where they are based, as long as their systems or outputs are placed on the EU market or affect EU users.
Is a visible label like 'AI-generated' enough, or do I need a machine-readable marker?
A visible label alone is not enough for the Article 50(2) marking duty. You must use a machine-readable format so the content can be detected as artificially generated or manipulated. A multi-layered approach combining metadata and invisible watermarks is recommended.
What is the difference between the Article 50(2) marking duty and the Article 50(4) disclosure duty?
Article 50(2) requires providers of generative AI systems to mark synthetic audio, image, video and text output in a machine-readable format. Article 50(4) requires deployers to disclose deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest. The marking duty is about technical detectability, while the disclosure duty is about informing users.
Does the editorial review exemption mean AI-generated articles do not need labelling?
AI-generated text under human review with editorial responsibility is exempt from the public-interest text disclosure in Article 50(4). However, if the text is published to inform the public on matters of public interest and is not under such review, it must be disclosed. The exemption does not apply to the Article 50(2) marking duty for providers.