Article 50(2) · provider duty · in force since 2 August 2026

Machine-readable marking of AI content

Article 50(2) requires providers of generative AI to mark synthetic audio, image, video and text in a machine-readable way, through watermarks or provenance metadata that the chain can read. That duty has applied since 2 August 2026. Only systems already on the market before that date get until 2 December 2026. Embed AI maps your marking, your suppliers and your contracts.

Marking track

2 weeks

From generative output to demonstrable marking

Role determination: are you the provider of this system or not

Technical test on watermark, metadata and robustness

Supplier clauses that enforce the marking

Evidence file for the supervisory authority and for your buyers

The only transition period ends on 2 December 2026

Article 50 has applied since 2 August 2026 and was not postponed. Regulation (EU) 2026/1744 moved high-risk Annex III to 2 December 2027 and Annex I to 2 August 2028, but left Article 50 untouched. There is exactly one exception: providers of systems placed on the market before 2 August 2026 have until 2 December 2026 for the marking duty in paragraph 2. Organisations starting only now are not preparing, they are late and have a short window left to put it right.

What does the marking duty in Article 50(2) actually require?

Paragraph 2 is a provider duty. Anyone placing a generative system on the market, or embedding one in a product under their own name, must mark the output so a machine can establish that the content was artificially generated or manipulated. Embed AI tests that across four layers.

The marking lives in the output, not in the interface

A visible label on your website is not a machine-readable marking. It is a signal in or alongside the file itself, such as a watermark or provenance metadata, that survives when the content is shared onward.

The marking must be reliable and robust

The solution has to withstand normal reuse such as recompression, cropping or a format conversion. A signal that disappears at the first edit does not meet the effectiveness requirement.

The duty follows the role, not the logo

If you embed a third-party model in your own product and put your name on it, you can become the provider yourself. Embed AI records per system who in your chain is provider and who is deployer.

The chain must be closed contractually

Point 12 of the Commission guidelines requires proportionate measures in distribution chains, including contractual arrangements. Without a clause with your supplier you have no grip on a duty that can still land on you.

Mind the cut-off: the transition period until 2 December 2026 applies only to paragraph 2 and only to systems placed on the market before 2 August 2026. For anything you place on the market after that date the marking duty applies immediately. Deployer duties under paragraphs 3 and 4 have no such transition.

Does the provider marking replace the deployer labelling duty?

No, and this is the costliest misconception Embed AI encounters in practice. Provider and deployer carry two separate duties that sit alongside each other.

Point 117 of the guidelines is explicit

A deployer may not rely on the machine-readable marking of the provider. Anyone publishing a deepfake or certain AI text of public interest must make that recognisable to people under paragraph 4 themselves.

Two layers, two audiences

Paragraph 2 targets machines and detection tools. Paragraph 4 targets the reader or viewer. A correct watermark therefore solves nothing for your marketing or communications channel.

Freelancers do not shift the duty

Point 14 states that a legal person remains the deployer, even when the work is in fact carried out by freelancers. Outsourcing does not move the obligation to the agency.

Advertising does not qualify for the creative exception

Point 122 provides that where the character is mixed, the informative or commercial character prevails. A campaign image with an artistic layer therefore remains advertising and falls outside the lighter modality.

When do you need this track?

This track is built for two groups: organisations that offer or embed generative AI themselves, and buyers who must enforce the marking with their supplier.

You ship a generative feature yourself

Your product generates text, image, audio or video for customers. You are then the provider and the paragraph 2 marking duty rests directly on you.

You embed a third-party model

You put your own name on a feature running on an external model. Embed AI determines whether that makes you the provider and which marking you then have to deliver.

You buy in and must enforce it

Procurement, IT or legal must record that the supplier marks output, how that is evidenced and what happens if the signal is lost.

What does Embed AI deliver?

Role determination per system: provider, deployer or both

Technical test on watermark, provenance metadata and robustness of the signal

Transition test: which system falls under 2 December 2026 and which does not

Implementation path for marking in your generation and publication pipeline

Contract clauses that lock the marking with your supplier, in line with point 12

Separation between the paragraph 2 marking and the paragraph 4 labelling duty in your channels

Evidence file to demonstrate how you comply, also without signing the code of practice

Action list with owner, priority and date per open duty

How does the track run?

1

Role determination and scope

Embed AI inventories which systems generate synthetic output and establishes per system whether you are provider, deployer, or both at once.

2

Technical test of the marking

We assess the existing watermark or provenance metadata on detectability and robustness, and test what survives recompression or conversion.

3

Supply chain test

We test what your model supplier actually marks, what evidence exists for it and which contractual arrangements are still missing.

4

Implementation and texts

You receive the implementation path for the marking plus the separate labelling texts for the channels covered by paragraph 4.

5

Evidence and handover

We record the evidence file and hand the action list to owners across product, legal and communications.

What happens if the marking is missing?

For a breach of Article 50 the supervisory authority can impose a fine of up to 15 million euro or 3 percent of worldwide annual turnover, whichever is higher. That is a ceiling, not an automatic outcome. The code of practice on transparency of AI-generated content of 10 June 2026 is voluntary to sign. Point 148 states that those who do not sign must demonstrate through other appropriate means how they comply, and point 149 that adherence can weigh as a mitigating circumstance in a fine. That makes your evidence file the core of this track. Point 124 of the Commission guidelines adds that the lighter regime never justifies an infringement of intellectual property or data protection rights.

What does it cost?

Embed AI works with fixed amounts, so you know where you stand up front.

Governance scan

EUR 2,950

Compact scan on role determination, marking and supply chain risk. Deductible from a follow-up track.

Readiness Sprint

EUR 9,900

Full readiness on AI systems, roles, duties and evidence, with the paragraph 2 marking track included.

Bundle

EUR 21,900

Readiness plus implementation support and assurance of the duties across your chain and channels.

Who does this work for?

Product and engineering

Teams that must build the marking into the generation and export pipeline and keep the signal robust.

Procurement and vendor management

Teams that must enforce marking with their model supplier and want to see the evidence periodically.

Legal, privacy and compliance

Teams that must keep paragraph 2 and paragraph 4 duties apart and defend the evidence file.

Marketing and communications

Teams that publish and therefore carry their own paragraph 4 labelling duty, separate from what the provider marks.

Afterwards you know

Per system whether you are the provider

Whether your marking is machine-readable and robust

Which system falls under 2 December 2026

Which clause your supplier must sign

How you can demonstrate compliance

Frequently asked questions

What is a machine-readable marking under Article 50(2)?

It is a signal in or alongside the output that lets a machine establish that the content was artificially generated or manipulated, for example a watermark or provenance metadata. The solution must be effective and robust, so it has to keep working through normal reuse such as recompression or a format conversion. A visible label in your interface is not a machine-readable marking.

Who must mark: the provider or the deployer?

The paragraph 2 marking duty rests on the provider of the generative system. The deployer has its own duty under paragraph 4 to make deepfakes and certain AI text of public interest recognisable. Point 117 of Commission guidelines C(2026) 5054 final is explicit: a deployer may not rely on the machine-readable marking of the provider.

Until when does the transition period run?

Until 2 December 2026, and only for paragraph 2 and only for systems placed on the market before 2 August 2026. For systems you place on the market after that, the marking duty applies immediately. Deployer duties under paragraphs 3 and 4 have no transition period.

Do I become the provider if I embed an external model in my product?

You can. If you put your own name or brand on a generative feature running on a third-party model, you may be considered the provider and the marking duty rests on you. Embed AI records per system who in your chain is the provider, so you do not rely on a role you do not actually hold.

What can the supervisory authority impose if the marking is missing?

For Article 50 the maximum is 15 million euro or 3 percent of worldwide annual turnover, whichever is higher. That is a ceiling, not an automatic outcome. Point 149 of the code of practice of 10 June 2026 states that adherence can weigh as a mitigating circumstance, and point 148 that those who do not sign must demonstrate compliance through other appropriate means.

Is this legal advice?

No. This is a practical implementation track. Embed AI structures roles, marking, contracts and evidence so product, procurement, legal and communications can act in a focused way. For formal legal advice, involve legal counsel.

Get your marking ready before 2 December 2026.

Start with the free transparency scan. Embed AI then determines per system whether you are the provider, whether your marking holds up and which arrangements your supplier still has to lock down.

Rivium Westlaan 46, Capelle aan den IJsselCoC 90283597