A company that uses an AI tool to generate text, images or audio tends to ask itself the same question when reading about Article 50: do I have to mark what I produce?
The short answer is no, or not through that route. The long answer has three parts that rarely appear together.
Who holds the obligation
Article 50(2) of Regulation (EU) 2024/1689 opens by naming the subject without ambiguity: "Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked…".
Providers. That is, whoever develops the system and places it on the market under its own name. If your company uses a third party's image-generation tool, the technical marking obligation falls on whoever builds it, not on you for using it.
This doesn't mean the deployer has nothing to do under Article 50. It does, but under other paragraphs and with different content — and confusing one with another is the source of a good part of the inflated task lists in circulation.
What the marking actually consists of
The second common misunderstanding is picturing a visible label, a stamp or a footer notice. The article says something else: the outputs must be "marked in a machine-readable format and detectable as artificially generated or manipulated".
Machine-readable. Metadata, technical watermarks, embedded signals: something a system can read, not necessarily something the reader sees. It's an infrastructure obligation, which is why it makes sense for it to fall on whoever builds the system.
The article also qualifies it: providers shall ensure their technical solutions are effective, interoperable, robust and reliable "as far as this is technically feasible", taking into account the limitations of each type of content, implementation costs and the state of the art. It isn't a guaranteed outcome: it's a duty to use the appropriate means to the extent the technology allows.
And it carries an explicit exception that's almost never cited: the obligation "shall not apply to the extent the AI systems perform an assistive function for standard editing". A spell-checker, or an assistant that reorders paragraphs or adjusts tone, doesn't fall under this just because it uses AI.
The transitional period until December 2026, and what it covers
This is the part most often misquoted, and for good reason: the date exists, but it doesn't mean what's usually said.
Regulation (EU) 2026/1744 — the Digital Omnibus, published in the Official Journal on 24 July 2026 — added a new paragraph to Article 111. It says that providers of systems generating synthetic content "that have been placed on the market before 2 August 2026" shall take the necessary steps to comply with Article 50(2) "by 2 December 2026".
Read it carefully, because every piece narrows the scope:
- It's a transitional arrangement, not a general date. 2 December 2026 isn't "when marking comes into force". It's the deadline for adapting what already existed.
- It only covers what was placed on the market before 2 August 2026. A generative system placed on the market after that date has no transitional period: Article 50(2) applies to it from the start.
- And it belongs to providers. The paragraph names them. A deploying company has no deadline of its own to use up here.
So the claim "marking isn't mandatory until December 2026" is false for every system launched from August onward — which is most of what's being adopted right now.
Why this error resembles the others
It's the same mechanism we've seen with the Annex III date and, in its most extreme form, with the Article 111(2) grace period, where the date disappeared from the article: a claim that was accurate at some point, or for some scenario, and that gets repeated outside its context without anything in the sentence giving warning.
Here the claim was never general: it was born narrowed to a specific scenario and spread without that narrowing. It's the hardest variant to correct, because whoever repeats it can point to a real text that backs it up — only it backs up less than what it's being made to say.
What to check if your company generates content with AI
Four checks, and none of them involve marking files by hand.
First: which obligation is actually yours. If you're the deployer and not the provider, Article 50(2) doesn't impose the technical marking on you. What you do have are the duties that are directed at you — and it's worth keeping them separate, because Article 50's transparency obligations are split across different subjects, and mixing them up produces lists nobody can complete.
Second: whether the provider marks, and how. It's a question of diligence, not an obligation of result on your part: it's either in their documentation or it isn't. Noting it down when you contract takes a minute; finding it out two years later, when someone asks, takes considerably longer. It's part of what's worth asking an AI vendor before signing.
Third: when the system you use was placed on the market. It determines whether the transitional period covered it. It's, once again, a piece of data nobody notes down when buying, and that's only missed when it's needed.
Fourth: whether your case falls under the standard-editing exception. Not every use of AI on a text generates synthetic content within the meaning of the article.
Where this leaves a mid-sized company
None of the four checks is difficult on its own. What makes them unmanageable is their shape: they're pieces of information known at different moments — when contracting, when deploying, when renewing —, known by different people, and only needed together the day someone asks.
A spreadsheet can hold them. What it can't do is answer, two years later, who noted each one down and when it was last reviewed — which is exactly what separates having the data from being able to prove it.
Dates and content in accordance with Articles 50 and 111 of Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744 (Official Journal of the EU, 24 July 2026, in force since 27 July 2026).
This article is for informational purposes only and does not constitute legal advice.