EU AI Act · Deadline 2 August 2026
On Sunday, Article 50 of the AI Act becomes applicable. For many companies this is the first part of the EU AI Act that becomes immediately visible: not in a documentation folder, but in their own chatbot, on their website, in their ad campaign.
The debate of recent weeks has been dominated by the Digital Omnibus, leaving many with the impression that the AI Act has been postponed. That is true for the high-risk obligations. It is not true for Article 50.
Applies from
2 August 2026
Transparency obligations under Art. 50: chatbot disclosure, deepfake labelling, emotion recognition, AI-generated text on matters of public interest.
Postponed to
2 December 2027
High-risk obligations under Annex III. For AI embedded in regulated products under Annex I: 2 August 2028.
What actually applies on 2 August
Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It amends more than 40 articles of the AI Act and reorganises the timeline.
Chapter IV, which contains Article 50, was not moved to a later stage. The general date of application under Article 113 remains 2 August 2026.
The timeline after the Omnibus
2 Aug 2026
Art. 50(1), (3) and (4) apply. Fines become possible.
2 Dec 2026
End of the transition period for Art. 50(2) for legacy systems.
2 Dec 2027
High-risk obligations under Annex III.
2 Aug 2028
AI embedded in products under Annex I.
In short: What was postponed are the extensive documentation duties for high-risk systems. What was not postponed is the duty to tell people that they are talking to an AI or looking at AI-generated content.
The four scenarios
Article 50 does not apply to every AI system. It attaches to four specific situations. What also matters is whether your company acts as the provider or the deployer in each case.
Interaction with people
Users must be informed that they are communicating with an AI system. This covers chatbots, voice agents and virtual assistants.
Machine-readable marking
Synthetic output in text, image, audio and video must be marked in a machine-readable format and be detectable as artificially generated.
Emotion recognition and biometrics
People exposed to emotion recognition or biometric categorisation must be informed. The GDPR applies in parallel.
Deepfakes and text
Deepfakes must be disclosed as artificially generated. The same applies to published AI-generated text on matters of public interest without editorial review.
Article 50(5) sets out how the information must be given: clearly, in a distinguishable manner, and at the latest at the time of the first interaction or exposure. The disclosures must meet accessibility requirements.
What changes in practice
Chatbots and voice agents in customer service
The disclosure has to appear where the interaction begins, not in the terms and conditions, not in hidden metadata and not in the URL. The Commission's final guidelines describe a multimodal approach: the notice must be perceivable through the same channel the interaction runs on. For a voice agent that means audible, not just in the legal notice.
Marketing with AI-generated visuals, avatars and voices
As soon as content depicts people, objects or events in a way that appears authentic without being real, the deepfake disclosure applies. The classic cases: the photorealistic AI model in a campaign, or the cloned voice in an ad. For artistic, satirical or fictional work, the disclosure must be designed so that it does not impair the display of the work. It does not disappear.
Editorial content and advice pages
The case that is most often underestimated. The Commission reads the concept of public interest broadly: politics, health, education, the environment, the economy and public security all fall within it. An AI-written health guide on a company website may therefore be covered. Plain product descriptions in a shop, internal communication and creative copy without informational intent generally are not.
Emotion recognition in customer contact
Anyone using sentiment analysis in call centres, recruitment interviews or retail spaces must inform the people affected. This is where Article 50 meets the GDPR directly, because special categories of personal data are regularly involved.
Four exceptions worth knowing
Obviousness
If it is obvious to a reasonably well-informed person that they are interacting with an AI, the disclosure under paragraph 1 falls away. Do not rely on it, the burden of proof sits with you.
Editorial responsibility
For AI-generated text, the labelling duty only falls away where there is genuine, substantial human review and a named party holding editorial responsibility. The guidelines read this narrowly. A quick glance before publishing is not enough.
Assistive function
Where the AI does not substantially alter the input data or their meaning, paragraph 2 does not apply. Spell checking yes, full text generation no.
No retroactive effect
Content created before 2 August 2026 does not have to be labelled retrospectively.
There is a transition period, but a narrow one
The Digital Omnibus created a transition period until 2 December 2026 for the machine-readable marking under paragraph 2. It applies exclusively to AI systems placed on the market or put into service before 2 August 2026. Systems launched after that date must comply immediately.
Common misconception
This period is addressed to providers and covers paragraph 2 only. It is not a general grace period. The obligations under paragraphs 1, 3 and 4, precisely the deployer duties for chatbot disclosure, emotion recognition and deepfakes, apply from 2 August 2026 with no transition.
The uncertainty is over
For a long time, the strongest argument against acting early was that the requirements were still too vague. That argument no longer holds.
On 10 June 2026 the Commission published the Code of Practice on Transparency of AI-Generated Content. It has a provider section and a deployer section and is formally voluntary. The Commission and the AI Board have nevertheless confirmed it as a suitable instrument for demonstrating compliance with the transparency obligations. Anyone who does not sign has to produce the same evidence by their own route.
On 20 July 2026 the Commission's final guidelines on Article 50 followed. They set out scope, exceptions and practical implementation, including the labelling of legacy content, AI-generated backgrounds in advertising, and systems that interact with children, where age-appropriate and accessible disclosures are required.
Who enforces this, and what does a breach cost?
€15 million
or 3 % of total worldwide annual turnover, whichever is higher (Art. 99(4)(g)).
The lower figure
applies to SMEs and start-ups (Art. 99(6)).
BNetzA
becomes Germany's central market surveillance authority where no sectoral authority is responsible.
The German Bundestag adopted the national AI implementation act on 11 June 2026. The Federal Network Agency (Bundesnetzagentur) will host a coordination and competence centre as well as a central contact and complaints office. BaFin remains in the lead for financial services, the BSI is involved for security-relevant systems, and the existing sectoral market surveillance authorities stay responsible for products under Annex I.
Realistically, no one will have an inspection at the door on Monday morning. The real lever is a different one: a complaints office means competitors, consumer protection bodies and affected individuals have an official route to report unlabelled systems.
Five steps for the coming week
Map every system with outside contact
Which AI systems produce content that goes out to the public, or interact directly with people? Chatbot, voice agent, image generator in marketing, text generator in editorial, analytics tools in customer service.
Establish your role for each system
Provider or deployer? A chatbot you build yourself makes you the provider, one you buy in makes you the deployer. The obligations differ, and both can apply in the same company at the same time.
Put the disclosure in the right place
Visible, perceivable and no later than the first interaction. Not in the terms and conditions, not in the footer.
Define your editorial process
If you want to rely on the editorial exception for text, you need a documented review process and a named responsible person. Without both, the exception will not hold up in a dispute.
Document the evidence
The obligation is met when you can prove it is met. Record which system falls under which scenario, what labelling is in place, who is responsible and when it was last reviewed.
The point that matters
Article 50 is not the most demanding obligation in the AI Act. But it is the first one that is visible from the outside. A missing disclosure in a chatbot is noticed by any customer who looks for it. Incomplete technical documentation is only noticed in an audit.
That is exactly why it pays not to treat 2 August as an isolated task. Mapping which AI systems run in your company, which role you hold for each of them and who is responsible creates the foundation that the high-risk obligations due by December 2027 will build on. The inventory you compile this week for Article 50 is the same one you will need for everything that follows.
With SimpleAct
Map your AI systems before someone asks about them.
EU AI Act and GDPR in one platform: register AI systems centrally, classify them rule-based, assign responsibilities and export audit-ready evidence at any time. Made in Germany, hosted in Germany.
Start your AI inventory →This article is for general information only and does not constitute legal advice. Last updated: 31 July 2026.
Tags
Yannick | SimpleAct Team
Author · SimpleAct Team