Most coverage of the Digital Omnibus on AI focused on the headline change: the Annex III high-risk deadline moving from 2 August 2026 to 2 December 2027. A narrower, easy-to-miss change affects a different obligation entirely: labelling AI-generated content and deepfakes under Article 50(2). If your product generates images, audio, video, or text that could be mistaken for human-made content, this is the deadline that actually applies to you, and it is not the one most people assume.
What Article 50(2) requires
Article 50(2) requires providers of AI systems that generate synthetic audio, image, video or text content to mark the output as artificially generated or manipulated, in a machine-readable format, detectable as such. In practice this covers image generators, voice cloning and text-to-speech tools, video synthesis, and deepfake generation, watermarking or embedded metadata, not just a visible label.
This sits alongside the broader Article 50 transparency layer we've covered before for chatbots: telling users they are talking to AI, and labelling deepfakes for viewers, both of which remain due on the original 2 August 2026 date.
What actually moved, and what didn't
The Digital Omnibus extended the Article 50(2) watermarking deadline specifically for systems already on the market at 2 August 2026, from 2 August 2026 to 2 December 2026, a four-month grace period. This extension exists because retrofitting machine-readable watermarking into an already-deployed generation pipeline is a real engineering task, not a configuration flag, and the original timeline gave existing products almost no runway.
Two things did not move:
- Any system launched after 2 August 2026 must implement Article 50(2) watermarking from day one. There is no grace period for new products; the extension is retroactive protection for legacy systems, not a general delay.
- The user-facing disclosure obligations, telling a person they are interacting with AI, and labelling visible deepfake content for the viewer, are unaffected. Those are due 2 August 2026 regardless of when the system launched.
So a content-generation product that has been live since 2025 gets until 2 December 2026 to add machine-readable watermarking to its output, but should already, as of August, be disclosing to users that its output is AI-generated where that disclosure applies.
Why this is easy to get wrong
The Annex III delay to December 2027 was widely reported and easy to conflate with "the AI Act obligations were pushed back." Article 50 was never part of that delay, and now has two internal dates of its own depending on when your system launched. A team that reads "the AI Act deadline moved to 2027" and stops tracking Article 50 entirely is the most common mistake we're seeing, and it is the one most likely to draw an actual information request, since Article 50 (unlike Annex III) is already fully enforceable today.
What to check now
- Was your generation feature live before 2 August 2026? If yes, your watermarking deadline is 2 December 2026. If it launched after, watermarking should already be in place.
- Is your user-facing "this is AI-generated" disclosure already live? This was never delayed and should be in production now.
- Does your watermark meet the "machine-readable, detectable" bar, not just a visible caption a user could crop out or a platform could strip?
Our free AI disclosure generator produces Article 50-compliant notices in 8 languages, including the IPTC metadata snippet relevant to image and video watermarking. Run EuroComply's free EU AI Act assessment to confirm which Article 50 obligations apply to your specific product and by which date.