The 23 September 2026 angle on the EU AI Act is not a surprise new duty: it is a countdown. Article 50 transparency has been in force since 2 August 2026; the grace period for machine-readable synthetic marking on generative systems already placed on the market or put into service before that date runs until 2 December 2026.
Per Commission FAQs and analyses from firms such as Cooley and Stibbe, providers must make synthetic audio, image, video and text machine-readably detectable. New systems since 2 August had to comply immediately; only the pre-August stock enjoys the window through December.
The provider-versus-deployer split matters. Article 50’s core loads onto whoever places the generative system on the market; deployers have duties to inform natural persons in interactions (chatbots) and to label deepfakes and certain public-interest texts under specific conditions. Mixing the two roles is September’s most common compliance error.
The Code of Practice on transparency of AI-generated content points toward layered marking—signed metadata plus imperceptible watermarks—without replacing the legal text. Teams waiting for the code to “clarify everything” in November risk hitting 2 December without a production watermarking pipeline.
The Digital Omnibus deferred elements of Annex III high-risk duties toward December 2027, but not Article 50’s core. That asymmetry breeds false ease: legal teams hear “deferrals” and drop their guard on transparency—the duty already live.
For generative providers with EU users, December pressure is industrial: integrate detectability without breaking latency, agree with hosts and CDNs on metadata preservation, and document which outputs are excepted (for example minor assistive content or artistic contexts per article nuances). Every misapplied exception is a potential fine.
Startups that fine-tune frontier models and expose them as EU SaaS are also on the map. You need not be OpenAI to be a “provider” for transparency purposes. The AI Office has already sent RFIs to more than 30 actors; Article 50 will be a standard question in any autumn-winter inspection.
Neural Pulse links the December grace to the rest of the 23 September agenda: international mandatory tests, geopolitical hotlines and cheap-model shipping. More synthetic tokens in circulation make marking more urgent, not less.
Technical risks are familiar: watermarks that die under recompression; metadata stripped by WhatsApp or social networks; detector false positives. Lawmakers ask for reasonable detectability, not magic; courts and the AI Office will set the bar through cases.
Q4 operating advice: inventory pre- and post-2 August systems; a marking plan by modality; signal-survival tests; and a single provider/deployer owner per product. Without that inventory, 2 December becomes improvisation.
Media and platforms should prepare labeling UX for deepfakes and public-interest texts aligned with deployer duties. Machine-readable marking grace does not defer every user-visible transparency obligation.
In short, as of 23 September 2026 the Article 50 clock shows under ten weeks for pre-August generative legacy systems to add machine-readable synthetic marking, while the transparency core already applies and the Omnibus did not freeze it.
We will track AI Office guidance, any early symbolic enforcement, and whether the Code of Practice becomes a de facto safe harbor or an insufficient floor.
Anyone treating December as “another deferral” has misread the calendar: the deferral was Annex III, not Article 50.
