AI Act Article 50: Labelling Rules for Chatbots, Voice Agents and AI Avatars

AI Act Article 50 has applied since 2 August 2026. Chatbots, voice agents and AI avatars that talk to people in the EU must disclose they are AI from the first interaction, unless obvious. Deepfakes must carry a visible label. Providers of generative systems launched before August have until 2 December 2026 to add machine-readable marks.

For most mid-sized and large companies, the real questions are which parts apply, who owns each part, and what must change on the website, in the contact centre and in marketing.

What AI Act Article 50 requires

Article 50 of the AI Act contains four separate duties:

1. Interaction disclosure: AI systems that interact directly with people must be designed so that people know they are dealing with AI, unless this is obvious to a reasonably well-informed person.
2. Machine-readable marking: Systems that generate synthetic audio, images, video or text must mark their outputs in a machine-readable, detectable way.
3. Emotion recognition and biometric categorisation: Deployers must inform the people exposed to these systems.
4. Deepfakes and public-interest text: Deployers must disclose deepfakes, and must label AI-generated text published to inform the public on matters of public interest, unless it has gone through human review or editorial control.

According to the European Commission's Article 50 FAQ, the information must be given clearly, at the latest at the first interaction or exposure, and in line with accessibility requirements. Fines for breaching these rules can reach €15 million or 3% of worldwide annual turnover, with proportionality for SMEs and small mid-caps.

Provider or deployer: which role does your company have?

A provider is an organisation that develops an AI system, or has it developed, and puts it into service under its own name or trademark. That second part matters. If an agency built a customer service bot for you on top of a large language model and it runs under your brand, your company is likely the provider of that system, not just a user of it.

A deployer is an organisation that uses an AI system under its authority in a professional context. Employees and contractors acting under its control are not separate deployers.

In practice:

• Off-the-shelf chatbot or voice widget: the vendor carries the design duty, but you still need to check that the disclosure is switched on and visible in your configuration.
• Bot or agent built for you and run under your brand: treat the interaction disclosure as your obligation.
• Marketing team publishing AI video or images: your company is the deployer and owns the deepfake label.

Chatbots and voice agents: say it at the first interaction

The Commission's guidance sets four cumulative criteria for the interaction duty: it must be an AI system, designed for a genuine two-way exchange, communicating directly (not through a human intermediary), with natural persons. Background systems and machine-to-machine processes fall outside it.

For a website chat assistant, a short line in the first message ("You are chatting with an AI assistant") is the simplest approach. For a voice agent on the phone, the disclosure must be audible at the start of the call. The "obvious" exception is read restrictively, so a human name and a friendly avatar make it harder to rely on.

Often missed: internal assistants used by employees, and AI agents that message customers on the company's behalf.

AI avatars and synthetic video: when a label is needed

A deepfake under the AI Act is AI-generated or manipulated image, audio or video that resembles existing persons, objects, places or events and would falsely appear authentic. All three elements (resemblance, something that exists or plausibly could, and a false appearance of authenticity) must be present.

That gives a practical test for AI avatars. A synthetic presenter that clearly looks like an animated character may not meet the definition. A realistic video of your CEO, created with an avatar platform such as HeyGen, delivering a message they never recorded on camera is much closer to the line. The Commission is explicit that deployers cannot rely only on the invisible machine-readable mark added by the provider: the disclosure must be perceivable by people, for example a visible or audible label.

AI-generated text and the editorial review exception

Most corporate blog posts, product pages and sales emails are not "public interest" text under Article 50(4). The rule targets published text informing the public on matters such as politics, public health, consumer safety or financial developments.

If your company publishes that kind of content, there is an exemption when the text has gone through real human review or editorial control and a person holds editorial responsibility. The Commission's FAQ states that spell-checking or grammar correction does not count. A named editor who checks the substance, the facts and the sources does.

Why 2 December 2026 matters

Article 50 has applied in full since 2 August 2026 to all systems in scope. There is one limited transition: generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking and detection duty. Content generated before 2 August does not need to be labelled retroactively, although the Commission encourages it.

The Commission has also backed a voluntary Code of Practice on transparency of AI-generated content, along with a set of EU icons that deployers can use as labels. Signing it is one way to demonstrate compliance with the marking and labelling duties; non-signatories must show equivalent measures. When enforcement started, the Commission published a first list of more than 180 signatories.

If you buy generative tools, ask each vendor in writing how and when their outputs are marked.

Who enforces Article 50 in Bulgaria?

Enforcement sits mainly with national market surveillance authorities, with the AI Office covering systems built on general-purpose models by the same provider. Bulgaria is behind on this. The Ministry responsible told BTA that market surveillance arrangements were still at draft stage, and the national AI bill submitted to Parliament on 24 September 2026 is still collecting opinions before first reading.

That is not a safe harbour. The regulation applies directly, and a regulator can be designated at any time.

A practical checklist for the next 30 days

1. List every AI system that faces people: website chat, voice agents, email agents, internal assistants, avatar videos, AI-generated images in campaigns.
2. Assign a role to each one: provider or deployer, using the "under your own name" test.
3. Check the first interaction: is the AI disclosure visible or audible from the first message or second of the call?
4. Set a labelling rule for synthetic media: when an EU icon or caption is required, and who approves it.
5. Ask vendors about marking: which standard they use and whether they meet the 2 December 2026 date.
6. Name an owner: one person accountable for AI transparency, linked to your wider AI governance work.

A quick way to see where the gaps are is the free AI Act readiness test from Encorp.ai, which flags the areas to address first.

Key takeaways

• Article 50 has applied since 2 August 2026; the 2 December 2026 date only covers marking for generative systems already on the market.
• A bot built for you and run under your brand can make your company the provider.
• Chatbots, voice agents and internal assistants need a disclosure at the first interaction.
• Realistic avatar videos of real people need a visible or audible label, not just an invisible mark.
• Bulgaria has no designated market surveillance authority yet, but the rules apply directly.