Article 50 of the EU AI Act has applied since 2 August 2026. One part of it has a second date. Under Article 50(2), providers of generative AI systems must mark what those systems produce in a machine-readable format, so that it can be detected as artificially generated. For systems that were already on the EU market before 2 August, the Commission gives a limited grace period: in the words of its Article 50 FAQ, providers of such systems "must comply with those obligations only as from 2 December 2026".
That is four weeks away. The reading we keep meeting goes like this: almost every organisation's AI stack predates August, so almost every organisation must now mark its AI content. The first half is true. The second half sends work to the wrong desk.
What the grace period covers, precisely
One obligation. The extension applies only to the marking and detection duty in Article 50(2). Telling people they are talking to an AI system under Article 50(1), and the deployer duties in 50(3) and 50(4), have applied since 2 August with no extension.
Existing systems only. It applies to systems placed on the market before 2 August 2026. A generative system launched since then has owed marking from the day it launched.
Not retroactive. The Commission says content generated before 2 August 2026 does not have to be labelled after the fact. It encourages deployers to do so where they can.
Not optional after the date. Article 50 failures carry fines of up to 15 million euro or 3% of worldwide annual turnover, whichever is higher (for SMEs, whichever is lower).
Whose deadline it is
Article 50(2) binds providers. A provider, in Article 3(3), is whoever develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark. The obligation travels with that role, not with the model inside the system.
That splits most organisations' exposure in two.
Where you use a supplier's assistant, the marking is theirs. If your staff use a chat assistant or an image tool sold by a model lab, the lab is the provider of that system and the 2 December duty sits with it. Your job is to know its position: ask the supplier, in writing, how its outputs are marked and how they can be detected, and keep the answer.
Where you built it, the marking may be yours. A customer-facing assistant, a document generator or an image feature that you built on a model API and run under your own name can make you the provider of that AI system, whoever trained the model underneath. Putting a system into service covers supply for your own use, so an internal tool you built is not automatically out of scope. If that system generates text, images, audio or video, the marking is your engineering problem, and the grace period only helps if it was in service before 2 August.
This is a practical reading, not legal advice, and counsel should confirm which role you hold for each system. The point is that the question has to be asked system by system. "We use Anthropic and OpenAI" answers it for none of them.
What machine-readable marking means in practice
The Act asks for marks that are effective, reliable, robust and interoperable, as far as is technically feasible. Recital 133 names the families of technique: watermarks, metadata, cryptographic methods for proving provenance and authenticity, logging methods and fingerprints. The Commission's guidelines add that the specifics and limits of each content type, the cost of implementation and the state of the art all count. Text is harder to mark durably than an image, and the guidelines do not pretend otherwise.
The Code of Practice on Transparency of AI-generated Content is the route the Commission and the AI Board have assessed as adequate. Providers can sign it and rely on its measures, or show compliance by alternative adequate means. The FAQ is candid about the second route: expect more requests for information.
The guidelines also narrow the scope. Short strings of numbers, symbols or letters fall outside it, as do machine-to-machine outputs never exposed to a person, closed-loop industrial and production outputs that are not the final output, and assistive standard editing. A narrow exemption for business-to-business and industrial contexts applies on conditions the guidelines set out. Each exemption is a claim you may later have to support with evidence.
Marking is an engineering task. Knowing which systems you must mark, and being able to show why the others are exempt, is a governance task. Only the second can be done in four weeks without a release.
The deployer side does not wait for December
A provider's machine-readable mark does not discharge a deployer's duties. The Commission is explicit that deployers "cannot simply rely on the machine-readable marking embedded in the content by the provider" to disclose a deepfake. Disclosure has to be perceivable by people without technical tools, a visible or audible label for example. And AI-generated text published to inform the public on matters of public interest needs a label unless it has had human review or editorial control and a person holds editorial responsibility for it, the exemption we read as a specification in The Article 50 exemption is the specification. Both have applied since 2 August.
Four weeks, four steps
1. List every generative system and its role. For each one: what it produces, who it reaches, whether you are the provider or the deployer, and the date it was first placed on the market or put into service. That date decides whether the grace period applies at all, so keep the evidence for it.
2. Get each supplier's marking position in writing. Where the supplier is the provider, its answer is your evidence. Where it cannot give one, that is a finding in its own right.
3. Mark what you provide. For systems where you are the provider, choose the technique per content type, test that the mark survives the routes your content travels, and decide between the Code of Practice and an alternative you are prepared to argue.
4. Write down every exemption you rely on. If an output is machine-to-machine, or business-to-business on the guidelines' conditions, record why, and who decided.
None of this is new work for an organisation that already knows where its AI-generated content comes from. As we argued in Your AI governance architecture is your Art.50 strategy, marking is the visible end of a provenance problem. We help organisations map which generative systems they provide and which they deploy, build the marking and the evidence for the ones they own, and put the record in place that answers a regulator's question from one source. If 2 December is on your calendar and you are not yet sure which systems it applies to, talk to us now.