Article 50(4) · deepfakes · in force since 2 August 2026

Deepfake compliance under Article 50

If you publish realistic synthetic images, audio or video, the Article 50(4) duty sits with you as deployer, not with your vendor. Embed AI sets up your deepfake compliance: we determine per publication channel which disclosure is required, record the exceptions and prepare the evidence. The duty has applied since 2 August 2026.

Deepfake compliance

1 to 2 weeks

From synthetic material to demonstrable disclosure

Inventory of realistic synthetic image, audio and video

Per publication: covered by paragraph 4 or by an exception

Disclosure texts and placement per channel

Work instruction and evidence file per publication channel

The duty applies now, not later

Article 50 has applied since 2 August 2026 and was not postponed. Regulation (EU) 2026/1744, in force since 27 July 2026, moved high-risk Annex III to 2 December 2027 and Annex I to 2 August 2028, but left Article 50 untouched. Organizations that publish realistic synthetic material are therefore late, not preparing. The only transition period sits in paragraph 2, applies solely to systems placed on the market before 2 August 2026 and runs until 2 December 2026. For your publications under paragraph 4 that room does not exist.

Why is deepfake compliance your problem and not your vendor’s?

Embed AI maps the role split per publication. The four duties of Article 50 do not sit with the same party, and paragraph 4 points directly at you as the publishing organization.

Paragraph 1: notice on direct AI interaction

The provider of a system that interacts directly with natural persons informs them that they are communicating with AI, unless this is obvious from the circumstances.

Paragraph 2: machine-readable marking

The provider marks synthetic output in machine-readable form. Only systems placed on the market before 2 August 2026 have a transition period until 2 December 2026.

Paragraph 3: emotion recognition and biometric categorisation

The deployer informs the exposed persons when it uses emotion-recognition or biometric-categorisation systems.

Paragraph 4: deepfakes and AI text of public interest

The deployer discloses deepfakes and certain AI-generated text on matters of public interest. If you publish the material, this duty is yours.

Point 117 of Commission guidelines C(2026) 5054 final of 20 July 2026 is decisive here: a deployer may not rely on the provider’s machine-readable marking. The watermark or metadata from your generation tool therefore does not satisfy your own duty under paragraph 4. You must arrange visible and comprehensible disclosure yourself, at the moment of publication.

When do you need this service?

This service is for organizations that publish realistic synthetic material and want to record when disclosure is mandatory and when an exception actually holds.

1

You publish synthetic image, audio or video

Campaign imagery, avatars, voice-overs, generated presenters or dubbed video that realistically resemble a real person, object or event.

2

You work with agencies and freelancers

Point 14 of the guidelines keeps the legal person the deployer, also when freelancers or agencies produce the material. Outsourcing does not move the duty.

3

You lean on the creative exception

Point 122 holds that where the character is mixed, the informative or commercial character prevails. Advertising therefore does not reach the creative exception.

What exactly do the Commission guidelines say?

Embed AI uses guidelines C(2026) 5054 final of 20 July 2026 as the assessment framework. Four points drive almost every decision on synthetic material.

Point 117

A deployer may not rely on the provider’s machine-readable marking. Your own disclosure is a standalone duty.

Point 12

Distribution chains require proportionate measures, including contractual arrangements with parties that create or distribute material for you.

Point 122

Where the character is mixed, the informative or commercial character prevails. A creative form does not rescue an advertisement.

Point 124

The lighter regime of Article 50 does not justify any infringement of intellectual property or data protection rights.

Do you need to sign the code of practice?

The code of practice on transparency of AI-generated content was published on 10 June 2026 and signing is voluntary. Point 148 makes clear that anyone who does not sign must demonstrate through other appropriate means how they comply. Point 149 adds that compliance may count as a mitigating circumstance when a fine is set. Embed AI helps you make that choice on solid ground and delivers the evidence you need either way.

What the service delivers

Inventory of all realistic synthetic publications per channel and per brand

Assessment per publication: deepfake under paragraph 4, AI text of public interest, or out of scope

Assessment of the creative, satirical and artistic exception per item

Ready-to-use disclosure texts with placement, timing and styling per channel

Work instruction for marketing, communications and editorial, including an approval step

Contract clauses for agencies, freelancers and vendors in line with point 12

Evidence file: who published what, with which disclosure and on which ground

Position on the code of practice on transparency of AI-generated content

Approach in 1 to 2 weeks

1

Scope and intake

We determine which channels, brands and production routes produce realistic synthetic material and who owns them.

2

Inventory of publications

We map per channel which image, audio and video is generated or manipulated and how realistic the depiction is.

3

Assessment against paragraph 4

We test every item against the deepfake definition and the exceptions, and record the reasoning per item.

4

Disclosure and supply chain

We deliver the texts and placement per channel, plus contractual arrangements for agencies and vendors in line with point 12.

5

Evidence and embedding

You get a work instruction, an approval step in the publication process and a file that shows how you comply.

Who this works for

Marketing and brand

Teams that use generative AI for campaign imagery, avatars or voices and need to know when a label is mandatory.

Communications and press

Teams that publish on matters of public interest and want a sharp line between editorial text and AI text.

Legal, privacy and compliance

Teams that want to translate the paragraph 4 duties into testable criteria, contracts and evidence.

Studio, video and production

Teams using synthetic voices, dubbing or generated imagery that must build disclosure into the workflow.

What happens if you arrange nothing?

For 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 practical risk usually sits earlier, in reputation and in a publication stream you cannot reconstruct: without a file you cannot show on which ground an item went out without a label.

What does it cost?

Embed AI works with fixed amounts, excluding VAT. You choose the entry point that fits your situation.

Governance scan

EUR 2,950

Compact scan of your synthetic publications and role split. Deductible from a follow-up engagement.

AI Act Readiness Sprint

EUR 9,900

Full readiness on AI systems, roles, duties and evidence, with deepfake compliance included.

Bundle

EUR 21,900

Readiness plus implementation and embedding, including work instructions, contracts and a training track.

Afterwards you know

Which publications fall under paragraph 4

Why your vendor’s watermark is not enough

When the creative exception actually holds

Which arrangements your agencies need contractually

How you demonstrate compliance per item

Frequently asked questions

Who must disclose a deepfake: we or our AI vendor?

You, if you publish the material. Article 50(4) places the duty on the deployer. Point 117 of Commission guidelines C(2026) 5054 final of 20 July 2026 states explicitly that a deployer may not rely on the provider’s machine-readable marking. The watermark or metadata from your generation tool is therefore no substitute for your own visible disclosure.

Does our AI campaign fall under the creative exception?

Almost never. The exception for evidently artistic, creative, satirical or fictional work is narrow and interpreted strictly. Point 122 holds that where the character is mixed, the informative or commercial character prevails. An advertisement in a creative form therefore does not reach the exception. Even where the exception does apply, an appropriate mention remains necessary that does not hamper the display of the work.

We outsource production to an agency or freelancer. Does that remove our duty?

No. Point 14 of the guidelines keeps the legal person the deployer, also when freelancers are used. Point 12 additionally requires proportionate measures in distribution chains, including contractual arrangements. Embed AI supplies the clauses and the work instruction to record this in your supply chain agreements.

Is there still a transition period for deepfakes?

No. The only transition period in Article 50 sits in paragraph 2, applies only to systems placed on the market before 2 August 2026 and runs until 2 December 2026. It covers machine-readable marking by the provider, not your duty as deployer under paragraph 4. For your publications the duty has applied since 2 August 2026.

What is at stake if we do not comply?

For 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. Point 149 of the guidelines names compliance with the code of practice on transparency of AI-generated content as a possible mitigating circumstance when a fine is set.

Should we sign the code of practice?

That is voluntary. The code of practice on transparency of AI-generated content was published on 10 June 2026. Point 148 states that anyone who does not sign must demonstrate through other appropriate means how they comply. Embed AI helps you weigh the choice and makes sure you hold a file that carries that demonstration either way.

Arrange your deepfake compliance before the next publication goes live.

Start with the free transparency scan. You immediately see whether disclosure on your synthetic media holds up. Embed AI then determines per publication channel which disclosure is mandatory and which evidence you record.

Rivium Westlaan 46, Capelle aan den IJsselCoC 90283597