In this article9 sections
  1. Article 50 Is Now Binding
  2. What Article 50 Actually Says
  3. The Chatbot Disclosure in Practice
  4. Deepfake Labelling: The Detail That Catches Companies Out
  5. Who Bears Responsibility: Providers vs. Deployers
  6. Penalties for Non-Compliance
  7. A Practical Disclosure Checklist for Deployers
  8. The Single Most Urgent Action
  9. References

Article 50 Is Now Binding

AI Act Article 50 transparency obligations apply from 2 August 2026. Providers of generative AI systems already on the market before that date have until 2 December 2026 for machine-readable marking (Art 50(2), Art 111(4)). If your company runs a customer service chatbot, uses an AI voice agent, or publishes AI-generated images or video, these rules apply to you now.

The rule is straightforward in principle: people interacting with an AI system have a right to know they are doing so. But the details of Article 50 matter enormously for how you design disclosures, draft policies, and train your team.

What Article 50 Actually Says

Article 50 sets out four distinct transparency obligations. Each one targets a different scenario.

Obligation 1: Chatbot self-disclosure. Providers must design AI systems that interact directly with people so those people are told they are talking to an AI, unless that is obvious to a reasonably well-informed person (Art 50(1)). The information must be clear and given at the latest at the first interaction (Art 50(5)), not buried in a terms-of-service page. As a deployer, do not switch that notice off.

Obligation 2: Emotion recognition and biometric categorisation. Deployers of a lawful emotion recognition or biometric categorisation system must inform the people exposed to it (Art 50(3)). Note: emotion recognition at work (except for medical or safety reasons) and biometric categorisation that infers, for example, political opinions or sexual orientation are prohibited outright under Article 5(1)(f) and (g). An HR tool that "reads" candidate sentiment is not a transparency question: it is banned.

Obligation 3: Synthetic content marking and deep fake disclosure. Providers of generative tools must mark their output in a machine-readable way (Art 50(2)). If you as deployer publish a deep fake (realistic content of real-looking people, places or events that could pass as authentic), you must disclose that it is AI-generated or manipulated (Art 50(4)). Think of synthetic voice in marketing calls, AI-generated photos of people in advertising, or video edited by generative AI.

Obligation 4: AI-generated text on matters of public interest. Text generated by AI and published to inform people on topics such as elections, health, or financial matters must be labelled as AI-generated, unless a human has reviewed it and someone holds editorial responsibility for the publication. News publishers, healthcare communicators, and financial services firms should pay close attention here.

The Chatbot Disclosure in Practice

Picture a visitor landing on your website and starting a chat. Your AI customer service agent responds within milliseconds. Under Article 50, that first response, or an automated pre-chat message, must make it clear an AI is responding. The phrase does not have to be dramatic. Something as simple as "You are chatting with an AI assistant" is enough, as long as it is prominent and intelligible.

One carve-out exists: no notice is needed where it is obvious, to a reasonably well-informed, observant and circumspect person, that they are dealing with an AI (Art 50(1)). Treat that threshold as high. A human name and a profile picture on an automated agent point the other way.

Companies building on a model API such as ChatGPT for their own customer-facing chatbot are usually the provider of that chatbot, so the disclosure is theirs to build in. The model vendor does not do this for you.

Deepfake Labelling: The Detail That Catches Companies Out

The synthetic content labelling requirement is where many SMEs underestimate their exposure. Consider a few scenarios that are already common across European businesses:

  • A marketing team uses an AI image generator to create photos of diverse "customers" for a campaign.

  • A training department produces AI-narrated video modules with a synthetic presenter.

  • A financial services firm uses a personalised AI voice to call customers about account changes.

The generator's provider must mark this output in a machine-readable format (Art 50(2)). Where the content shows realistic people or events that could pass as real, you as deployer must also disclose that it is AI-generated (Art 50(4)). For video and audio, the marking usually means watermarking or metadata. The Commission published guidelines on Article 50 (20 July 2026) and a voluntary Code of Practice on marking and labelling AI-generated content.

For AI-generated images, the C2PA (Coalition for Content Provenance and Authenticity) standard is emerging as the practical implementation route. Tools like Adobe Firefly already embed this metadata. If you are using other generators, check whether they offer equivalent tagging and, if not, apply your own metadata before publishing.

For evidently artistic, satirical or fictional work, the disclosure is lighter: you still disclose that AI-generated content is present, in a way that does not spoil the work (Art 50(4)).

Who Bears Responsibility: Providers vs. Deployers

This is the question most compliance officers ask first. The short answer: both, and the obligations are layered.

Providers of AI systems must build the disclosure and marking into their products (Art 50(1) and (2)). Article 50 itself places separate duties on deployers (Art 50(3) and (4)). And if the provider supplies a chatbot platform with a disclosure banner and you switch it off to make the interface look cleaner, the gap is yours.

For SMEs, this has a very practical implication. When you procure any AI customer service, HR, or communications tool, your vendor contract should confirm:

  1. Which Article 50 disclosure functions are built into the platform.

  2. What customisation you are responsible for configuring.

  3. Whether the tool supports machine-readable labelling for synthetic content outputs.

If your vendor cannot answer those three questions, treat that as a compliance risk to document.

Penalties for Non-Compliance

Under Article 99 of the AI Act, breaching Article 50 can be fined up to €15 million or 3% of total annual worldwide turnover, whichever is higher; for SMEs and start-ups the lower of the two applies (Art 99(4)(g), 99(6)). National authorities set the actual penalty.

National market surveillance authorities supervise Article 50. In the Netherlands the Autoriteit Persoonsgegevens is the intended supervisor for transparency under the draft Dutch implementing law; it has no AI Act powers until that law is in force.

A Practical Disclosure Checklist for Deployers

Here is what to review before your next audit or vendor onboarding:

  • Map every AI touchpoint where your organisation communicates with customers, employees, or the public. Include chatbots, voice agents, email automation with generative text, and social content.

  • Confirm chatbot disclosure timing. The disclosure must appear at or before the first AI-generated message, not after several exchanges.

  • Audit synthetic content outputs. For every AI tool that produces images, video, or audio, confirm whether machine-readable labelling is enabled and what standard it uses.

  • Review vendor contracts. Add Article 50 capability confirmations to your standard software procurement checklist.

  • Train your content and customer service teams. Staff need to understand when a disclosure is required and how to escalate if a tool does not support labelling. This connects directly to the Article 4 AI literacy obligation, which has applied since 2 February 2025.

  • Document your decisions. If you determine that a specific use case falls under an exemption, write that reasoning down. If an authority questions your practices, documented decision-making is your strongest defence.

The Single Most Urgent Action

Pull up your customer-facing AI tools today. Open each one and test whether a new user receives an AI disclosure before or at the moment of the first interaction. If the answer is no, or if you are unsure, that is your starting point.

Transparency under Article 50 does not require expensive system overhauls for most SMEs. It requires deliberate configuration, clear vendor conversations, and a written record of what you have done and why. Check, configure, document.

Run a free 2-minute compliance check at comply.khairos.ai to see where your organisation stands on Article 50 and the other obligations that apply to you.

References

  1. Article 50

  2. Article 99

  3. Autoriteit Persoonsgegevens: next step in AI supervision (9 Oct 2026)

  4. Code of Practice on AI-generated content

  5. Article 4

  6. comply.khairos.ai

Part of our guide to High-Risk & Prohibited AI

Not sure where your company stands?

Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.

Am I compliant? Check in 2 minHow Khairos Comply works