Article 50(2): Marking Is Half the Rule, Detection Is the Other Half

Expert contribution

By Andrei Brisc

Most of what is written about Article 50(2) of the EU AI Act is about watermarks. That is half of the rule. The Article asks for two things: that AI output is marked in a machine-readable format, and that it is detectable as AI-generated. The second half is the one most teams have not built yet, and from 2 December 2026 it also applies to systems that were already on the market.

This is a general explanation, not legal advice on any specific product.

What Article 50(2) asks for

Providers of AI systems that generate synthetic audio, image, video or text “shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated” (Article 50(2), Regulation (EU) 2024/1689). The technical solution has to be “effective, interoperable, robust and reliable as far as this is technically feasible” (Article 50(2), Regulation (EU) 2024/1689).

Marking is something you put into the output. Detection is something you give to everyone else, so they can check it. The Commission’s guidelines are direct: “the provider is obliged to ensure that the means of detection are available to the persons potentially exposed to the content” (Commission Guidelines on Article 50, paragraph 75).

A watermark that only you can read does not meet that. The guidelines expect detection that allows “any third party to implement detection”, with a provider’s own solution accepted only if it stays interoperable (Commission Guidelines on Article 50, paragraph 76).

Being small is no exemption. The guidelines call technical feasibility “an objective notion that is not dependent on the specific resources and capabilities of individual providers” (Commission Guidelines on Article 50, paragraph 81).

Who counts as a provider

A provider is anyone who 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, whether for payment or free of charge” (Article 3(3), Regulation (EU) 2024/1689).

A company that takes an existing generative AI system, modifies it and puts it into service under its own name “becomes a provider of the new system” (Commission Guidelines on Article 50, paragraph 11). Using someone else’s model does not change this: a “downstream provider” is a provider of an AI system that integrates a model “provided by another entity” (Article 3(68), Regulation (EU) 2024/1689). Chatbot builders, AI content tools and agencies with a branded product should assume the duty may sit with them.

“Article 50 AI Act does not explicitly apply to GPAI models”, although model providers are encouraged to mark content at model level (Commission Guidelines on Article 50, paragraph 27).

So you can build on the model provider’s marking. You still carry the proof. Providers “may rely on the marking solution implemented by an upstream model provider” if it is compliant, but that reliance “is without prejudice to the responsibility of the provider of the AI system to demonstrate compliance” (Commission Guidelines on Article 50, paragraph 74).

The Regulation applies to providers outside the EU where the output of their system is used in the Union (Article 2(1)(c), Regulation (EU) 2024/1689).

So you can build on the model provider’s marking. You still carry the proof.

What falls outside the rule

The duty does not apply where the system performs “an assistive function for standard editing” or does “not substantially alter the input data provided by the deployer or the semantics thereof” (Article 50(2), Regulation (EU) 2024/1689).

The guidelines give examples on both sides (Commission Guidelines on Article 50, paragraph 92):

• Exempt: grammar correction and spellchecking, translations, format conversions, minor cropping, transcriptions of conversations.
• Still to be marked: AI-generated summaries, rewriting that changes style, structure or meaning, inserting or removing objects or people in images and video, and realistic speech in a specific person’s voice.

Some outputs are out of scope altogether: short sequences such as single words, alt-text and UI labels, source code, and output exchanged only between machines without human exposure (Commission Guidelines on Article 50, paragraph 68).

The dates

Article 50 has applied since 2 August 2026 (Article 113, Regulation (EU) 2024/1689). A product launched after that date has had to comply from launch.

Systems already on the market before 2 August 2026 received a transition. Their providers “shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026” (Article 111(4), Regulation (EU) 2024/1689, inserted by Regulation (EU) 2026/1744).

For chatbots, a system that is partly interactive and partly generative gets the transition only for marking; the duty to tell people they are talking to an AI has applied since 2 August 2026 (Commission Guidelines on Article 50, paragraph 153). Content generated before 2 August 2026 does not need to be marked retroactively (Commission Guidelines on Article 50, paragraph 154).

Breaches of Article 50 can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher (Article 99(4)(g), Regulation (EU) 2024/1689). For SMEs, including start-ups, the lower of the two applies (Article 99(6), Regulation (EU) 2024/1689).

What marking and detection look like in practice

The Regulation does not prescribe a technique. Recital 133 lists several, from watermarks and metadata to fingerprints, alone or combined (Recital 133, Regulation (EU) 2024/1689).

The practical benchmark is the Code of Practice on Transparency of AI-Generated Content, published in final form on 10 June 2026 (Code of Practice, European Commission). It is voluntary. The guidelines describe adherence to a code assessed as adequate as “a straightforward, predictable, and legally certain way of demonstrating compliance” (Commission Guidelines on Article 50, paragraph 147). The Code itself adds that adherence “does not constitute conclusive evidence of compliance” (Code of Practice, Section 1, objective (a)).

For signatories, the Code asks for:

• Two layers of marking. In most cases, “no single marking technique suffices” (Code of Practice, Section 1, recital (b)). Signatories use digitally signed metadata plus an imperceptible watermark (Code of Practice, Sub-measures 1.1.1 and 1.1.2). Free-form text cannot carry metadata, so one watermark layer is enough there, and text longer than 200 tokens still needs it (Code of Practice, Measure 1.1 and Sub-measure 1.1.2). Fingerprinting or logging alone is not enough (Code of Practice, Sub-measure 1.1.3).
• A detection solution, free of charge. A public specification, software or a cloud API, with one detection mechanism per marking technique and results downloadable in a digitally signed format. A narrow fee option exists only for signatories with fewer than 1,000,000 monthly users, and only for very high request volumes (Code of Practice, Measure 2.1).
• Interoperable detection by 2 February 2027 (Code of Practice, Measure 3.4).

Two points are often misread. First, none of these texts names a specific standard such as C2PA. Open standards can help with interoperability, but no named one is a legal requirement. Second, providers who do not sign are expected to show compliance by other means, for example with “a gap analysis” against the Code’s measures, and “may also be subject to a larger number of requests for information” (Commission Guidelines on Article 50, paragraph 148).

The common mix-up: 50(2), 50(1) and 50(4)

Article 50 holds three duties that are often confused (Article 50, Regulation (EU) 2024/1689):

• 50(1): providers of systems that interact with people must make sure those people are told they are interacting with an AI system, unless that is obvious. This is a message to a human.
• 50(2): providers of generative systems must mark output in a machine-readable way and make it detectable. This is a signal to machines and checkers.
• 50(4): deployers must disclose deep fakes, and AI-generated text published to inform the public on matters of public interest, unless it has had human review and someone holds editorial responsibility. This is a visible label.

One does not cover another. A machine-readable mark that users cannot perceive does not satisfy 50(1), but it can still serve 50(2) (Commission Guidelines on Article 50, paragraph 38). The reverse also holds: a banner saying “You are interacting with an AI system” does not mark a single output.

A short self-check

1. Does the product generate text, audio, images or video that a person will see or hear, beyond single words, labels or code?
2. Is it released under your name or brand, even if the model comes from someone else?
3. Was it on the market before 2 August 2026? If so, your marking date is 2 December 2026. If not, it already applies.
4. Where does your marking come from: your own system, or the model provider? If upstream, can you show it meets Article 50(2)?
5. If someone outside your company receives a piece of your output today, can they check whether it is AI-generated, and get an answer?
6. Do your terms prohibit users from removing or tampering with metadata marks, as the Code asks signatories to do (Code of Practice, Measure 1.2)?
7. If the product talks to people, does it also tell them they are talking to an AI?

If question 5 is the one you cannot answer, start there. It is the half of the rule that anyone outside your company can test.

These questions help you find gaps. They do not replace an assessment of your specific system.

This article is an expert contribution. It reflects the author’s views and is provided for general information only. It is not legal advice. AI Governance Desk publishes expert contributions for educational purposes.