Article 50 has applied since 2 August 2026. The duties are allocated by actor and use case. Providers disclose direct AI interaction when it is not obvious and machine-mark synthetic output under paragraphs 1 and 2. Deployers inform on emotion recognition or biometric categorisation and disclose deepfakes plus certain public-interest text under paragraphs 3 and 4. Only providers of relevant paragraph 2 systems placed on the market before 2 August 2026 have until 2 December 2026 for machine-readable marking.
For many organisations this is the sharpest AI Act deadline with enforcement in 2026. Whereas the core high-risk obligations for Annex III systems apply from 2 December 2027, Article 50 stays on 2 August 2026. Regulation (EU) 2026/1744 was published on 24 July 2026 and entered into force on 27 July 2026. Below you will read what Article 50 asks, who must do what per role and which steps to take now.
What does Article 50 ask exactly?
Article 50 contains four specific transparency duties and is not a high-risk regime. It requires no conformity assessment, Annex IV technical documentation or EU-database registration. The precise disclosure, information duty or marking follows from actor and use case, not from a generic rule that all AI content needs a visible label.
The obligations fall into four parts, divided between providers and deployers. A provider develops the AI system or places it on the market under its own name. A deployer puts the system into use under its own responsibility. Many organisations are provider and deployer at the same time for different systems, and must therefore determine per system which duty falls to them.
Who must do what, per role?
The table below sets the four parts of Article 50 against the role that carries the obligation and the date it applies.
| Part | What | Role | Applies |
|---|---|---|---|
| 50(1) | Disclose a chatbot or AI interaction | Provider | 2 August 2026 |
| 50(2) | Machine-readable marking of generated audio, image, video, text | Provider | 2 August 2026, existing systems until 2 December 2026 |
| 50(3) | Inform people exposed to emotion recognition or biometric categorisation | Deployer | 2 August 2026 |
| 50(4) | Visibly mark deepfakes and certain AI text for public information | Deployer | 2 August 2026 |
Article 50(1) falls to the provider. A system intended to interact directly with people, such as a chatbot or virtual assistant, must be designed so that the person knows they are talking to AI. This is not required where that is obvious to a reasonably observant person.
Article 50(2) also falls to the provider and is technically the most demanding duty. Generative systems that produce audio, image, video or text must mark that output in a machine-readable format and make it detectable as artificially generated or manipulated. This touches on watermarking and provenance standards such as C2PA. An exception applies to systems with a purely assistive, editorial function, such as a spell checker.
Article 50(3) falls to the deployer. Anyone deploying a system for emotion recognition or biometric categorisation must inform the people exposed to it. Note that certain forms of emotion recognition are already prohibited as a banned practice since 2 February 2025, so check Article 5 first before relying on the transparency route.
Article 50(4) also falls to the deployer and has two branches. Anyone who generates or manipulates a deepfake must disclose that the content is artificial. For artistic, creative, satirical or fictional work a lighter form applies: the disclosure may not get in the way of enjoying the work. In addition, AI-generated text published to inform the public on matters of public interest must be marked as such, unless the text has been reviewed under human editorial responsibility.
An organisation may hold a different role for each system. Integrating or using an external model does not automatically make it the provider; Article 25 and the facts determine whether a role shift occurs. Contractually record how provider marking is delivered and preserved.
What is the role of watermarking and the transition period?
The machine-readable marking under Article 50(2) is the only duty with this separate transition. Providers of relevant systems placed on the EU market before 2 August 2026 have until 2 December 2026 to put that marking in order. For later systems, the provider duty applies immediately.
The other three parts of Article 50 have no transition period. Chatbot disclosure, informing people about emotion recognition or biometrics, and the marking of deepfakes and public AI text all have applied in full since 2 August 2026.
Which steps do you take now?
A practical preparation runs in four steps that you can complete well before the deadline.
Inventory which systems fall under Article 50
Map which systems interact directly with people, generate synthetic output, use emotion recognition or biometric categorisation, and which deepfakes or public-interest texts are published. Then determine whether paragraphs 1 to 4 actually apply to each situation.
Assign the role per system
Determine whether you are provider, deployer or both. The duties under 50(1) and 50(2) fall to the provider, those under 50(3) and 50(4) to the deployer. For external models, set out contractually who provides the marking.
Set up the disclosure and the marking technically
Add the chatbot disclosure, arrange the notice for biometrics or emotion recognition, and prepare the machine-readable marking via standards such as C2PA. This is the part that takes the most lead time, because watermarking and provenance signals are not in place overnight.
Build an evidence layer
Article 50 does not prescribe a formal conformity assessment. Screenshots, configurations, supplier information and internal policy can nevertheless help substantiate the disclosure or marking applied.
What happens in case of non-compliance?
Since 2 August 2026 Article 50 falls under the enforcement of the competent national supervisory authorities. In the Netherlands, supervision of the AI Act is being set up with the Dutch Data Protection Authority and the Dutch Authority for Digital Infrastructure in a coordinating role. For a breach of the transparency obligations a fine via the supervisory authority can follow, which under the regulation can run up to 15 million euro or 3 percent of total worldwide annual turnover, whichever is higher. That makes Article 50 the sharpest deadline with enforcement of 2026.
How do you put this into practice?
The legal explanation is one thing, the implementation another. For execution, Embed AI runs an Article 50 transparency check: a focused scan that determines the role per system, runs through the four parts, sets up the disclosure and marking, and orders the evidence layer. That connects to the broader AI governance scan and the Readiness Sprint in which transparency comes together with inventory, risk classification and governance.
The human side of transparency starts with awareness. Staff need to distinguish actor, use case and applicable duty. LearnWize records role-based learning measures, assessments and completions without guaranteeing compliance. For broader background, see the AI Act readiness roadmap.
Deeper reading per obligation
For each part of Article 50 there is a dedicated analysis:
- Labelling deepfakes (Article 50(4))
- Does your chatbot have to say it is AI? (Article 50(1))
- Machine-readable marking of AI content (Article 50(2))
- Emotion recognition and biometric categorisation (Article 50(3))
- AI-written text for public information (Article 50(4))
- Enforcement and fines for non-compliance
- Field test: do Dutch chatbots tell you they are AI?
Frequently asked questions about Article 50 transparency
Short, citable answers for organisations preparing for 2 August 2026.
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe