Article 50 EU AI Act: An Overview of the Transparency Obligations

Article 50 of the EU AI Regulation (Regulation (EU) 2024/1689) governs four transparency obligations for limited-risk AI systems: the disclosure of AI interaction, the machine-readable marking of synthetic content, notification in cases of emotion recognition or biometric categorisation, and the labelling of deepfakes and public-interest text. The substantive obligations apply from 2 August 2026. Article 50 sits between the unregulated minimal-risk tier and the high-risk regime and applies irrespective of any high-risk classification - many limited-risk systems are subject exclusively to the Art. 50 obligations.

The Three Most Important Answers in Brief

• Four scenarios, two responsible parties: Providers bear the obligations under para. 1 (AI interaction) and para. 2 (machine-readable marking); deployers bear the obligations under para. 3 (emotion recognition/biometric categorisation) and para. 4 (deepfakes and AI-generated texts relevant to the public).
• Key date 2 August 2026: Substantive application from 2 August 2026; the technical implementation deadline for marking and watermarking under para. 2 ends, following the political agreement on the Digital Omnibus, on 2 December 2026 (as of 2026, subject to formal adoption).
• Fine risk: Breaches fall under Tier 2 pursuant to Art. 99 - up to EUR 15 million or 3 percent of total worldwide annual turnover, whichever is higher.

The Four Transparency Scenarios in Detail

Article 50 addresses four clearly delineated situations. Each has its own trigger, its own responsible party and its own system of exemptions.

1. Disclosing AI Interaction (Art. 50(1))

Where an AI system is designed to interact directly with natural persons, the provider must design and develop the system in such a way that the persons concerned are informed that they are interacting with an AI. This obligation only ceases to apply where this is obvious from the circumstances and the context of use.

In practice, this primarily concerns customer service chatbots as well as voice and outbound calling bots. The "obvious from the context" exemption is interpreted narrowly: a human-sounding voice agent must disclose its AI nature. Where mid-market deployers procure a white-label chatbot, the Art. 25 trap is acute - by re-branding a system that has already been placed on the market, they switch into the provider role, and the para. 1 obligation becomes their own.

2. Marking Synthetic Content in a Machine-Readable Format (Art. 50(2))

Providers of AI systems that generate synthetic audio, image, video or text content - including general-purpose AI models used for content generation - must mark the outputs in a machine-readable format as artificially generated or manipulated. The technical solutions must be effective, interoperable, robust and reliable.

This obligation therefore falls on the model and tool level, not primarily on the end user. The Regulation does not prescribe any specific technical standard; it only defines the quality requirements (effective, interoperable, robust, reliable). The more detailed specification is provided through secondary instruments of the Commission (see the section on the Code of Practice and guidelines).

3. Disclosing Emotion Recognition and Biometric Categorisation (Art. 50(3))

Deployers of an emotion recognition system or a system for biometric categorisation must inform the natural persons exposed thereto about the operation of the system. Importantly: emotion recognition in the workplace and in educational institutions is, as a matter of principle, prohibited under Art. 5(1)(f) (with narrow medical and safety exemptions). The notification obligation of para. 3 therefore only applies to the permissible remaining cases outside these prohibited contexts.

4. Labelling Deepfakes and AI-Generated Texts Relevant to the Public (Art. 50(4))

Here the obligations lie with the deployer and break down into two sub-cases:

• Deepfakes (para. 4, first subparagraph): Anyone who generates or manipulates image, audio or video content that constitutes a deepfake must disclose that the content has been artificially generated or manipulated. In the case of artistic or satirical works, disclosure is limited to an appropriate, non-disruptive extent - the artistic exemption is interpreted narrowly.
• Texts on matters of public interest (para. 4, second subparagraph): Anyone who generates or manipulates AI-generated texts that are published in order to inform the public on matters of public interest must disclose that they have been artificially generated - unless the text has undergone a human editorial review or control and a natural or legal person bears the editorial responsibility.

Overview: Scenario, Obligation and Exemption

From When Does Article 50 Apply? Deadlines and the Digital Omnibus

The substantive transparency obligations of Art. 50 apply from 2 August 2026. For the technical implementation of the marking and watermarking solutions under para. 2, the originally six-month transitional period was shortened to three months by the political agreement on the Digital Omnibus of 7 May 2026; it ends on 2 December 2026.

Important compliance note (as of 2026, subject to change): While the Digital Omnibus does postpone the majority of the high-risk obligations (Annex III now from 2 December 2027, Annex I from 2 August 2028), it does not postpone the transparency obligations of Art. 50 as such. As long as the Digital Omnibus has not been formally adopted - the Parliament and Council have pledged adoption before 2 August 2026 - the original date of 2 August 2026 remains the legally authoritative default date. Those who plan their roadmap conservatively will adhere to this earlier deadline.

Machine-Readable Marking: Requirements and Secondary Instruments

Art. 50(2) requires a machine-readable marking that is effective, interoperable, robust and reliable - but does not prescribe any particular technical standard. The Regulation thus sets the result (a robust, interoperable marking) but leaves the concrete technical implementation to the provider and to further specification by the Commission.

This specification is provided through secondary instruments. The Commission published a first draft of a Code of Practice on the labelling of AI-generated content on 17 December 2025, a second draft on 3 March 2026, as well as draft guidelines on the implementation of the Art. 50 transparency obligations on 7-8 May 2026 (for consultation). These address the robustness of watermarks, the taxonomy (fully AI-generated vs. AI-assisted) and obligations relating to detection tools. Until these instruments are finalised, the legal benchmark remains the four-criteria test of the wording.

A Concrete Example With Figures

A DACH-region online retailer operates a procured customer service chatbot (8,000 conversations per month) and produces AI-generated product videos for a social media campaign. Three obligations apply:

• Chatbot (para. 1): The provider must have built in the AI disclosure technically. As the deployer, the retailer verifies that the notice is visible and not hidden.
• Product videos as deepfakes (para. 4): As the deployer, the retailer must label each generated clip as artificially generated.
• Machine-readable marking (para. 2): The generative model used must mark the outputs in a machine-readable format.

Were the retailer to market the chatbot under its own brand (re-branding under Art. 25(1)(a)), it would switch into the provider role and additionally assume the para. 1 obligation itself. If it fails to label the videos, it risks a Tier 2 fine: up to EUR 15 million or 3 percent of total worldwide annual turnover. Assuming an annual turnover of EUR 40 million, 3 percent would equal EUR 1.2 million; since the absolute maximum amount of EUR 15 million is higher, this applies as the upper limit for a non-SME. For SMEs and start-ups, however, the lower of the two amounts applies pursuant to Art. 99(6) - here, therefore, the EUR 1.2 million.

Practical Implementation: The Art. 50 Checklist

1. Inventory every customer-facing AI interaction point.
2. Confirm that the provider disclosure under para. 1 is operational and visible.
3. For each generative output type, check whether the machine-readable marking under para. 2 is present.
4. For each deepfake or AI-text use case, implement the deployer-side disclosure under para. 3/4.
5. Document the analysis wherever the "obvious from the context" or "human editorial review" exemption is claimed.
6. Update privacy notices, terms and conditions and AI usage notices accordingly - these often simultaneously fulfil the information obligation under Art. 26(11) for high-risk decisions pursuant to Annex III.

For Agencies and B2B Decision-Makers

Marketing agencies sit at a dual interface when it comes to Art. 50: they produce AI-generated content (paras. 2/4) and implement chatbots for clients (para. 1). The question of role - provider or deployer - determines who is liable, and it can shift quickly through re-branding under Art. 25. For B2B decision-makers the rule is: those who maintain a clear inventory of AI touchpoints, align machine-readable marking early on with an effective, interoperable and robust solution, and keep their exemption documentation clean are prepared for 2 August 2026 - and structurally reduce the Tier 2 fine risk.

Blck Alpaca supports DACH-region companies and agencies with exactly this assessment: from the touchpoint inventory through the clarification of roles to the labelling architecture.

Note: This article is intended for professional orientation and does not constitute legal advice. For a binding assessment of your specific case, please seek qualified legal counsel.