Article 50 Transparency Obligations Explained

On 2 August 2026, Article 50 of the EU AI Act became applicable. It is the Act's transparency layer. Its rules make sure people know when they are talking to a machine. They also make sure people know when content is AI-generated, and when a system reads their emotions or biometric traits. The Digital Omnibus deferred the high-risk regime's timeline, but it did not delay Article 50 — it applies now.

This guide walks through the four obligations and who each one falls on (provider or deployer). It also covers the exceptions, how and when you must give the information, and what non-compliance costs.

The Four Obligations at a Glance

Article 50 is titled "Transparency obligations for providers and deployers of certain AI systems". Each of its first four paragraphs targets a different situation:

50(1): Tell People They Are Talking to an AI

Providers must design AI systems that interact directly with natural persons so those persons know they are interacting with an AI system. The obligation falls away only where that is obvious. The test is the view of a reasonably well-informed, observant and circumspect person, given the circumstances and context of use. A branded support chatbot in a "Chat with our AI assistant" widget may pass that bar. An AI voice agent answering your phone line almost certainly does not.

A narrow carve-out covers systems authorised by law to detect, prevent, investigate or prosecute criminal offences. It does not apply where the system is available for the public to report an offence.

50(2): Mark AI-Generated Content at the Source

Providers of AI systems that generate synthetic audio, image, video or text — including general-purpose AI systems — must mark the outputs in a machine-readable format. The outputs must be detectable as artificially generated or manipulated. The marking must be effective, interoperable, robust and reliable as far as technically feasible. This takes into account the state of the art and relevant technical standards. In practice, that means provenance metadata and watermarking schemes.

The obligation does not apply where the system performs an assistive function for standard editing. It also does not apply where the system does not substantially alter the input data provided by the deployer, or its meaning. A further exception covers use authorised by law to detect, prevent, investigate or prosecute criminal offences.

Expect the detail to firm up. Article 50(7) tasks the AI Office with facilitating codes of practice on detection and labelling of artificial content. The Commission can approve those codes — or, if they fall short, set common implementation rules — via implementing acts.

50(3): Emotion Recognition and Biometric Categorisation

Deployers of an emotion recognition system or a biometric categorisation system must inform the natural persons exposed to it of the operation of the system. They must also process the personal data involved in line with the GDPR and the other applicable EU data-protection rules. The same law-enforcement carve-out as above applies, subject to safeguards.

Remember the interplay with the prohibitions. Since February 2025, Article 5 bans AI systems that infer emotions in the workplace and education institutions. The only exception is use for medical or safety reasons. Article 50(3) transparency only becomes relevant where the use is not prohibited in the first place.

50(4): Label Deep Fakes and AI-Written Public-Interest Text

Some AI systems generate or manipulate image, audio or video content that constitutes a deep fake. Their deployers must disclose that the content has been artificially generated or manipulated. For evidently artistic, creative, satirical, fictional or analogous works, the duty softens. Disclosure must happen in an appropriate manner that does not hamper the display or enjoyment of the work.

A second limb covers text. It covers deployers of an AI system that generates or manipulates text published with the purpose of informing the public on matters of public interest. They must disclose that the text is AI-generated or manipulated. The duty lapses if the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. That editorial-responsibility exception keeps a normal newsroom workflow with AI-assisted drafting out of scope.

How and When to Disclose

Article 50(5) sets the delivery standard. The information must reach the persons concerned in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure. It must also conform to applicable accessibility requirements. A notice buried in terms of service does not meet that bar.

Article 50(6) makes clear these duties stack. They do not replace the Chapter III high-risk requirements. They do not replace any other transparency obligations in EU or national law. A high-risk system that also interacts with people needs both sets of measures. GDPR information duties also continue to apply in parallel.

Why 2 August 2026 Matters

Article 113 makes 2 August 2026 the Act's general application date. Article 50 sits in that general tranche. The Digital Omnibus (provisional agreement, May 2026) deferred the high-risk timeline. It moved Annex III to 2 December 2027 and Annex I to 2 August 2028. It did not delay Article 50. The transparency obligations are live obligations today, not a future deadline.

Non-compliance with Article 50 carries administrative fines of up to EUR 15 000 000 or 3 % of total worldwide annual turnover (Article 99(4)(g)). The higher of the two amounts applies. Non-compliance is also an express ground for market-surveillance intervention against the system itself.

A Practical Compliance Sequence

For most organisations, Article 50 compliance is a screening exercise across the AI inventory:

1. Inventory — list every system that talks to people, generates content, reads emotions or biometric traits, or produces deep fakes. Include embedded vendor tools, not just what you built.
2. Map — match each system to the paragraph(s) that bind it. One system can trigger several: a customer-facing generative chatbot engages both 50(1) and 50(2).
3. Decide exemptions and write them down — obviousness, assistive editing, artistic works, editorial responsibility. Record the rationale so you can defend it later.
4. Implement — the disclosure at first interaction, machine-readable output marking, notices for emotion/biometric systems, deep-fake labels — in every deployment language, meeting accessibility requirements.
5. Keep evidence — screenshots of the notices, marking specifications, and a per-system disclosure record you can hand to a market-surveillance authority.