Article 50 explained. The transparency rules that hit on August 2, 2026

What Article 50 actually says

Article 50 is the EU AI Act's transparency obligation. It targets four AI uses and tells you that the affected people must know AI is involved. The wording in the official text is dense, but the four triggers are clear once you list them.

1. AI systems that interact with humans. Chatbots, voice agents, virtual assistants. The user must be told they are interacting with an AI, at the moment of first interaction, in a way that is appropriate to the context.
2. Synthetic content. AI-generated or AI-manipulated text, image, audio, or video. The provider of the system has to mark the output in a machine-readable way (watermark or metadata) so platforms can detect it. The deployer has to display the AI nature in a visible way for end users.
3. Deepfakes. AI-generated or manipulated images, audio, or video that closely resemble real people, objects, places, or events. The deepfake has to be clearly labelled as such, unless it is part of an evidently artistic or satirical work (and even then, the disclosure cannot interfere with the work itself).
4. Emotion recognition and biometric categorisation. People exposed to these systems must be informed about the operation of the system, with consent and processing under GDPR for any personal data involved. Note: the use of these systems is sometimes prohibited under Article 5 (workplaces, schools, sensitive characteristics).

What "clear, timely, accessible" means in practice

The text repeatedly uses three words to describe what counts as a valid disclosure: clear, timely, accessible.

Clear

The user must understand that AI is involved. Plain language. No technical jargon. No fine print. "Hey there, this site uses cutting-edge LLM technology" is not clear. "You are chatting with an AI assistant" is clear.

Timely

The disclosure must appear at the first interaction, not after the user has typed a question and got an answer. For a chatbot, this means the welcome message must contain the disclosure, or a banner must show before the input box appears. For a voice agent, the AI announcement is the very first sentence of the call.

Accessible

Users with disabilities must be able to perceive the disclosure. Screen readers must be able to read it. The disclosure should not be a pixel of small grey text on a grey background. Contrast, font size, ARIA labels.

The chatbot rule (paragraph 1)

"Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use."

The bar for "obvious" is high. A voice with a robotic tone does not always make AI nature obvious, because many real call centres use synthesised voices for menus. To be safe, assume the disclosure is required for every chatbot you operate, regardless of how AI-sounding it is.

The synthetic content rule (paragraph 2)

"Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."

This is two halves of the same obligation. The provider of the generator (OpenAI, Anthropic, Google, the image vendor) is on the hook for embedding a machine-readable marker. You as the deployer are on the hook for displaying a human-readable disclosure when you publish the output.

For images, the de facto standard is C2PA (Coalition for Content Provenance and Authenticity). Major generators (DALL-E, Adobe Firefly, Midjourney since v6) embed C2PA metadata automatically. For text, no equivalent watermark exists yet. The visible label is your responsibility either way.

The deepfake rule (paragraph 4)

"Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated."

A deepfake is any AI-generated content that closely resembles a real person, object, place, or event. The label has to be in a way that is clearly visible to users. For artistic, satirical, fictional or analogous work, the disclosure can be limited to a way that does not hamper the display, but it must still exist.

The emotion AI rule (paragraph 3)

"Deployers of an emotion recognition system or a biometric categorisation system shall inform the natural persons exposed thereto of the operation of the system and shall process the personal data in accordance with Regulation (EU) 2016/679, Regulation (EU) 2018/1725 and Directive (EU) 2016/680, as applicable."

You need to inform users in advance, get appropriate consent under GDPR for any biometric data, and never use these systems in the prohibited contexts (workplace, schools, sensitive characteristics).

What you must have ready by August 2, 2026

1. Every chatbot or voice agent shows an AI disclosure on first contact.
2. Every published AI-generated text, image, audio, or video has a visible label.
3. Every image generator your team uses checks for machine-readable provenance (C2PA), or you embed it yourself when the source generator does not.
4. Every deepfake (avatar, voice clone, AI face) is labelled as artificial.
5. Every emotion or biometric categorisation system informs the user and supports an opt out.
6. Screenshots of every live disclosure are saved as evidence.
7. A dated dossier exists, ready to hand to a regulator.

How ActHub maps to Article 50

Each Article 50 trigger has a matching tool inside ActHub.

• Chatbot disclosure rule. Use the Disclosure Widget with type "chatbot" and paste the JS snippet on your site.
• Synthetic content rule. Use the inline label snippet inside your blog post template and the AI Content Label widget.
• Deepfake rule. Use a Disclosure Widget with type "deepfake" near every AI avatar.
• Emotion AI rule. Use the type "emotion" widget plus the matching template in the Templates Library.
• Evidence collection. Use the Article 50 Checklist to track every item and the Compliance Dossier PDF to export the final audit pack.

The penalty math

Article 99 sets the fines. Non-compliance with Article 50 falls in the second tier: up to €15 million or 3 percent of total worldwide annual turnover, whichever is higher. SMEs and start-ups can have the lower of the two figures applied, per Article 99(6). For comparison, a complete chatbot disclosure setup with ActHub takes about 30 minutes and costs €19 per month on the Starter plan.

FAQ

Does Article 50 apply to me even if my chatbot is open source?

Yes. The Act regulates how the system is used, not where the code came from. If the chatbot interacts with humans in the EU, Article 50 applies.

Can I bury the disclosure inside my terms of service?

No. Article 50 requires the disclosure to be clear, timely, and accessible. Reading the terms of service is none of those things. The disclosure must appear at the moment of first interaction.

What if the chatbot is obviously a robot?

The Act allows an exception only when the AI nature is unmistakable for a reasonably well-informed person, given the circumstances. In practice, assume disclosure is required unless your chatbot has a robotic voice and visibly mechanical responses.

Do I need a separate disclosure for each AI feature?

Each disclosure obligation is tied to its trigger. A site with a chatbot, AI-written articles, and AI-generated images needs three disclosures (one per trigger), but you can place them appropriately in context.

Sources

• Article 50 full text
• Article 99 penalties
• C2PA content provenance standard
• Article 5 prohibited practices

This article is general information, not legal advice.