Article 50 of the EU AI Act became enforceable on 2 August 2026, and it splits the transparency work in two. The provider of a generative system embeds machine-readable marking in the output. The deployer puts a visible label on what gets published. If you build a product on somebody else's model, you are the deployer, and the visible label is yours.
The Code of Practice that around 190 organisations signed states plainly that no single marking technique satisfies all four qualities the law asks for, so the expected answer is layers, metadata plus a watermark, with no agreed benchmark for detecting either.
Systems already on the market before 2 August have until 2 December 2026 to meet the marking obligation. Four months is the whole runway, and it is an engineering runway, not a legal one.
What actually turned on
Article 50 covers four situations, and they do not all bind the same party.
Chatbot disclosure binds the provider. If a person is interacting with an AI system, they have to be told, unless it is already obvious.
Machine-readable marking binds the provider. Synthetic audio, image, video and text has to be marked in a machine-readable format and detectable as artificially generated. The regulation's own wording is that the marking must be effective, interoperable, robust and reliable, as far as technically feasible.
Emotion recognition and biometric categorisation notice binds the deployer. If you run one of those on people, you tell them.
Deepfake and public-interest text disclosure binds the deployer. Publish an AI-generated or manipulated image, audio or video of a person, or AI-generated text meant to inform the public on a matter of public interest, and you disclose it visibly.
Penalties reach 15 million euro or 3 percent of worldwide annual turnover, whichever is higher. They apply to providers, deployers, importers and distributors putting systems on the EU market, and to anyone whose output lands in front of EU users regardless of where the company sits.
Provider marks, deployer labels
The reading I keep encountering is that this is the model vendors' problem. OpenAI marks its images, Google watermarks with SynthID, Anthropic signed the code, so the box is ticked upstream and the application layer inherits it.
That is half right, and it is the wrong half.
The machine-readable marking obligation genuinely does sit with the provider. If you call an API and the returned image carries a C2PA manifest and an invisible watermark, that specific duty was discharged by someone else.
The visible label was not. Article 50(4) binds the deployer, meaning whoever uses the system in a professional capacity. That is you, in your product, at the point where content reaches a reader. A C2PA manifest is not a disclosure to a human being. It is signed metadata that survives until the first tool in your pipeline strips it, which is roughly the first resize.
The practical question for this week is not whether your model vendor is compliant. It is two narrower ones. Does your pipeline preserve the marking your vendor embedded, and does your interface tell a human being that this thing was generated.
The Code splits along exactly these lines. Section 1 is for providers of generative systems and for vendors of marking and detection technology. Section 2 is for deployers. Around 190 organisations signed by the end of July, 82 on Section 1 and 152 on Section 2. The Section 1 names read like a model release list: Anthropic, Google, Meta, Microsoft, Mistral, OpenAI, Cohere, Aleph Alpha, Black Forest Labs, Synthesia. Section 2 reads like the rest of the economy: Getty Images, Lenovo, Lufthansa, Bulgari.
The gap between 82 and 152 is the honest shape of this regulation. Far more companies are going to be labelling than marking.
The code concedes the technology is not there
The most useful sentence in the whole package is an admission.
The Code asks signatories to combine techniques, metadata, watermarking, provenance mechanisms, and to run more than one machine-readable layer where necessary. The stated reason is that no single technique currently meets all four legal requirements at once. Metadata is interoperable and trivially removable. Invisible watermarks survive processing better and do not interoperate across vendors. Post-hoc forensic detection is not considered reliable enough, and common evaluation benchmarks have not emerged.
Regulation arriving ahead of settled technology is normal. What is less common is a code of practice writing the gap down and then prescribing redundancy as the mitigation. Two imperfect layers, on the theory that their failure modes are uncorrelated.
For anyone building, that is a design instruction rather than a legal footnote. Do not pick a provenance standard and stop. Carry both layers, and treat marking as something you can lose at every hop. Two task forces launch in September for signatories to work implementation out among themselves, which is a fair signal of how settled any of this is.
Text is where the edges go soft
Text is the modality the framework handles least confidently, and the reasons are technical rather than political.
The deepfake definition does not extend to text. AI-generated text triggers disclosure only in the narrower case where it is published to inform the public on matters of public interest. Ordinary editing is carved out explicitly: spellchecking, grammar correction, quality improvements, format conversion. Text that went through substantive human editorial review, with a named person or organisation holding responsibility, is also out.
Those are reasonable carve-outs, and they are where most of the ambiguity now lives. Artistic, satirical and fictional works get a flexible disclosure regime with no detailed guidance on what flexible means. Whether a given edit counted as substantive review or as a quality improvement is a case-by-case call that enforcement practice will settle, not close reading.
If your product generates text and publishes it, the question to answer internally is narrow. Are we informing the public on a matter of public interest, and can we name who reviewed it. Write the answer down before somebody asks for it.
Four months, and what to do with them
The grandfathering clause is the part with a real date attached. Systems placed on the market before 2 August 2026 have until 2 December to satisfy the marking obligation. Content generated before 2 August needs no retroactive labelling.
Three things worth doing while that clock runs.
Audit what your pipeline destroys. Generate an asset from your provider, run it through the full chain, storage, resize, CDN, and check whether the C2PA manifest is still attached at the end. In most stacks it is not, and the loss is silent.
Decide where the visible label goes, and put it at first exposure rather than behind an info icon. The Code's language on deployer labelling is about visibility on first contact, not availability on inspection.
Write down your role per feature. Provider, deployer, or both, feature by feature. Most teams are both and have never separated the two, and the obligations genuinely differ.
Signing the Code is not the same as compliance, but the asymmetry is real. Signatories can point to it to demonstrate compliance. Non-signatories document their own equivalent measures under questioning. Either path needs you to know which of your features generate, which publish, and what survives in between.
The regulation that took effect this week does not ask you to solve provenance. It asks you to know where in your own stack it breaks.