Article 50 for Chatbots and AI Agents: What You Must Disclose

The chatbot labelling obligation derives from Article 50 of Regulation (EU) 2024/1689 (EU AI Act): users must be able to tell that they are interacting with an AI rather than a human. The disclosure must be made clearly and in good time, unless the AI nature is obvious from the circumstances. The transparency obligations apply from 2 August 2026. For agencies that ship chatbots and voice agents, this is the central obligation.

The three most important answers upfront

• Who must label? Article 50(1) addresses the provider, that is, the party who makes the system available under its own name. In white-label deployment, this role can pass to the agency (Article 25).
• When does the obligation apply? The transparency obligations apply from 2 August 2026; the technical deadline for machine-readable labelling ends on 2 December 2026 (as at 2026, subject to change).
• What does a breach cost? Up to 15 million euros or 3 percent of total worldwide annual turnover (fine tier 2 under Article 99).

What Article 50 specifically requires

Article 50 sits between the unregulated minimal-risk tier and the high-risk regime. It applies regardless of high-risk classification: many chatbots and voice agents are subject solely to these transparency obligations and to no further high-risk requirements. The Regulation recognises four triggers.

For classic chatbots and voice agents, Article 50(1) is decisive: AI systems intended to interact directly with natural persons must be designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI. The obligation falls on the provider and applies unless this is obvious from the circumstances and the context of use.

Also relevant for agents that produce content:

• Article 50(2) requires the provider to mark synthetically generated audio, image, video or text content as artificially generated or manipulated in a machine-readable format. This also affects GPAI models used for content generation.
• Article 50(3) requires the deployer of an emotion recognition or biometric categorisation system to inform the exposed natural persons of its operation.
• Article 50(4) requires the deployer to disclose deepfakes as well as AI-generated texts on matters of public interest, with narrowly interpreted exceptions for artistic works and for texts subject to human editorial responsibility.

When and how the notice must be given

The EU AI Act does not prescribe any fixed wording or rigid placement. What is required is that users are informed clearly and in good time, before or precisely when the interaction begins. Three principles help with implementation:

1. In good time means: at the start of the conversation, not only in the small print or after several messages. For voice agents, the notice must come audibly at the start of the conversation.
2. Clear and not hidden means: clearly visible, not buried in the terms and conditions. As the deployer of a third-party chatbot, you should check whether the provider's disclosure is actually present and not concealed.
3. Comprehensible means: recognisable to the average user without any technical prior knowledge.

Proven formulations for practice (not prescribed texts, but examples that fulfil the transparency objective):

• Chat notice: "You are chatting with an AI assistant."
• Bot greeting: "Hello, I am an AI-powered assistant. How can I help?"
• Voice agent at the start of the conversation: "You are speaking with an automated AI assistant."

The exception: obvious to recognise

Article 50(1) does not apply where the AI nature is obvious from the circumstances and the context of use. This exception is interpreted narrowly. A voice agent that sounds human must disclose its AI nature, precisely because it is not obvious. A chatbot with a human name and profile picture likewise does not fall under the exception. Anyone relying on it should document the justification in writing - in the event of a dispute, the responsible party bears the burden of argumentation.

Overview: who must label what?

Practical example with figures: service chatbot in retail

An Austrian online retailer has its agency build a customer service chatbot that runs on a GPAI model and handles around 8,000 conversations per month. The bot bears a human-sounding name and an avatar image. Classification: limited risk, Article 50(1) applies; because the bot makes no decisions about access to essential services, no high-risk classification arises.

If the agency ships the bot under the retailer's brand name, the retailer is regularly the provider and bears the Article 50(1) obligation. If, on the other hand, the agency rebrands it and distributes it as its own product, it becomes the provider itself under Article 25 and the obligation migrates to it. Implementation: a visible notice "You are chatting with an AI assistant" at the start of the conversation, plus a documented assessment that the "obvious" exception precisely does not apply here (human name plus avatar).

A worked example on the risk: in the absence of labelling, a tier 2 fine looms - up to 15 million euros or 3 percent of total worldwide annual turnover, whichever is higher. With a group turnover of 200 million euros, 3 percent would already be 6 million euros; here the percentage value would apply, not the absolute amount. For SMEs, start-ups and small mid-cap enterprises (SMCs, added by the Digital Omnibus agreement), the lower of the two values applies in each case under Article 99(6).

Deadlines and the Digital Omnibus caveat

The substantive transparency obligations of Article 50 apply from 2 August 2026. The technical transitional period for machine-readable labelling and watermarking was shortened from six to three months by the political Digital Omnibus agreement of 7 May 2026 and ends on 2 December 2026 (as at 2026, subject to formal adoption).

Important: the Digital Omnibus postpones the high-risk obligations (Annex III now from 2 December 2027), but not Article 50 in substance - there, only the technical implementation deadline is shortened. Until formal adoption - promised before 2 August 2026 - the original dates remain the legal standard. Alongside this, the EU 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, and draft guidelines on the Article 50 transparency obligations on 7 and 8 May 2026.

For agencies and B2B

Anyone who ships chatbots or voice agents should settle three things immediately: first, the clarification of roles in the contract - who is the provider, who is the deployer, and does white-label deployment tip the agency into the provider role under Article 25? Second, the disclosure component as a standard in every bot: a visible notice in the chat, an audible notice for the voice agent. Third, documentation of the exception assessment, should you rely on "obvious to recognise". Blck Alpaca supports DACH agencies and B2B teams in integrating this labelling cleanly, in a user-friendly and audit-proof manner, into existing chatbot stacks.

Note: This article serves as professional orientation and does not constitute legal advice. For a legally binding assessment of your specific use case, please consult qualified legal counsel.