The Article 50 marking deadline: what actually changes on 2 December 2026

I keep hearing the same relieved sentence from founders who shipped an AI feature earlier this year: "we're fine, we've got until December." Half of that is true. The other half is the part that gets a company fined, so it is worth separating them properly.

What actually happens on 2 December 2026?

The grace period for one specific piece of Article 50 runs out: machine-readable marking of AI-generated content, for systems that were already on the market before 2 August 2026. After that date, every generative AI feature you run, old or new, has to mark its own output in a way a machine can detect as artificially generated.

That is the whole deadline. It is not a deadline for telling people they are talking to an AI. It is not a deadline for labelling deepfakes. Those started on 2 August 2026, and nothing about the December date moved them, the same way the Annex III and Annex I high-risk deadlines moved but everything else in the Act did not.

Which of Article 50's four duties actually has a clock on it?

Article 50 bundles four separate obligations, and the confusion mostly comes from treating them as one thing:

• Bot disclosure. If someone is interacting with an AI system and it would not otherwise be obvious, you have to tell them.
• Biometric or emotion-recognition disclosure. If you deploy a system that recognises emotion or categorises people biometrically, the people exposed to it have to be informed.
• Deepfake and public-interest text labelling. If you deploy a system that generates deepfakes, or AI-generated text published on a matter of public interest, you have to disclose that it is artificially generated or manipulated.
• Content marking. If you provide a system that generates synthetic audio, image, video or text, the output itself has to carry a machine-readable mark, not just a disclosure to the person reading it.

All four became enforceable on 2 August 2026. The Digital Omnibus only carved out a transition window for the last one, and only for systems already live before that date: they get until 2 December 2026 to add the mark itself (Gibson Dunn, ComplianceHub.Wiki). The disclosure duty in that same feature, "tell the person they are talking to AI", was never on that clock. It applied from August, legacy system or not.

What counts as "shipped before 2 August 2026"?

A system you placed on the market or put into service before that date. If your chatbot, content generator, or AI writing assistant was live and in use before 2 August, it is the one that gets the extra runway on marking. Anything you ship for the first time after that date has to arrive with the mark already built in; there is no fresh three-month grace period for a feature you launch next week.

What does "machine-readable marking" actually require?

More than a footnote in the UI. The European Commission's Code of Practice on Transparency, finalised after a public draft and signed by roughly 190 organisations by the end of July 2026, points to a layered approach because no single technique is considered robust enough on its own: signed provenance metadata in the style of the C2PA standard for images, audio and video, plus an embedded watermark where the format supports one, with logging as a fallback where neither is technically feasible.

Text is the hard case, and it is worth knowing where your own vendor stands. Anthropic announced in August 2026 that every Claude model now weaves an imperceptible, machine-readable pattern directly into its generated text, applied worldwide rather than only for EU users, specifically as its answer to Article 50(2) (TechCrunch, Claude Help Center). If you built a feature on top of a model that already does this, a real chunk of your obligation is already handled upstream. If you built on a model that does not, or you post-process the output before it reaches a user, the mark can get stripped along the way and the obligation is back on you.

What if your AI feature is text-only, like a support chatbot?

Two different duties can apply to the same feature, and it is easy to satisfy one and miss the other. The bot-disclosure duty ("you are talking to an AI") is a UI problem: a line of text or a chat header solves it, and it has applied since August regardless of when you shipped. The content-marking duty is a data problem: the actual output has to carry something a machine can check, which a disclosure banner does not provide on its own. A chatbot that says "I'm an AI assistant" at the top of the window has done the first duty and nothing for the second.

What happens if you miss it?

Breaches of Article 50's transparency obligations sit in the AI Act's lower enforcement band: fines up to €15 million or 3 percent of global annual turnover, whichever is higher, with authorities directed to apply the lower of the two figures proportionately for SMEs and the newer small mid-cap category (Securing.AI). That is a real number for a small company, and enforcement discretion is not the same as being ignored.

Do you need a consultant, or an inventory?

Usually the second one, and it is smaller than it sounds. Most companies I talk to cannot immediately answer "which of our features generate synthetic content, and when did each one go live." That single list, features against launch dates, is what tells you whether you are racing the December deadline or already past it with room to spare. It is the same underlying work as AI governance and LLMOps: knowing what your AI systems actually do, with what data, before a regulator or a customer asks.

If you want a second opinion on which of your AI features need the mark and which just need a better disclosure line, book the free 30 minute intake call and bring the list. I will tell you honestly where you stand, including if the honest answer is "you already did this."