What is the AI transparency obligation?
The AI transparency obligation is set out in Article 50 of the AI Act, Regulation (EU) 2024/1689. Since 2 August 2026, anyone who communicates with an AI system has to know about it. If an AI agent answers customer questions via chat, WhatsApp or phone, that rule applies.
Article 50 contains four obligations:
- Article 50(1): AI systems that interact directly with people disclose that they are AI, unless this is already obvious.
- Article 50(2): systems that generate synthetic audio, images, video or text mark that output in a machine-readable format as AI-generated.
- Article 50(3): anyone who uses emotion recognition or biometric categorisation informs the people exposed to it.
- Article 50(4): anyone who creates deepfakes or publishes AI-generated text to inform the public on matters of public interest discloses this.
Under Article 50(5) you give this information clearly, accessibly and at the latest at the first interaction. This article follows the text of the AI Act, the European Commission guidelines of 20 July 2026 and its questions and answers on Article 50 (as of November 2026). It is general information; take your own situation to a lawyer. The full law is covered in the EU AI Act explainer.
Who must do what: provider or deployer?
The AI Act splits the obligations across two roles. The provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name. The deployer is the organisation that uses the system in a professional capacity.
If you buy an AI customer service tool from a software vendor, that vendor is the provider and you are the deployer. If you have your own AI agent built and put it into service under your own name, the definition in Article 3 can make you the provider.
| Obligation | Article | Who | What you actually do |
|---|---|---|---|
| Disclose that someone is interacting with AI | 50(1) | Provider | Build the system so it identifies itself as AI at the first interaction |
| Mark AI output in a machine-readable way | 50(2) | Provider | Watermark or metadata in generated images, audio, video and text |
| Disclose emotion recognition or biometric categorisation | 50(3) | Deployer | Inform people that such a system is operating |
| Disclose deepfakes | 50(4) | Deployer | A visible or audible notice with the image, video or audio |
| Disclose AI text on matters of public interest | 50(4) | Deployer | State that the text was made with AI, unless a human has reviewed it substantively and someone holds editorial responsibility |
Your customer sees your company name in the chat, so check for every tool yourself whether the AI disclosure is switched on.
When do you not have to disclose AI?
You do not have to disclose AI when it is obvious to a reasonably well-informed, observant and circumspect user that they are dealing with AI. According to the Commission, this exception must be interpreted restrictively. A robot icon in the chat header is therefore a risky basis.
According to the Commission, the obligation applies to a genuine two-way exchange in which a person communicates directly with the AI. An automatic confirmation from your contact form falls outside it. An AI agent that answers questions via chat, WhatsApp or phone falls within it. Callers often fail to recognise a voice agent with a natural voice as AI, so always disclose it there.
How do you disclose AI in customer service?
A good disclosure is short, sits in the first message and tells the customer how to reach a person. Article 50 does not require that last part; it follows from the position of the Dutch regulators below. Copy the lines and adapt them; the Dutch versions are for Dutch-speaking customers.
Website chat
Phone and voice agent
On the phone, the disclosure belongs in the opening sentence, before the first question, outbound calls included. More on voice agents in the article on an AI agent as a phone receptionist.
Must a customer always be able to reach a person?
Article 50 does not require human contact. The Dutch Authority for Consumers and Markets (ACM) and the Dutch Data Protection Authority (AP) do consider it part of the deal. In their joint opinion of 2 October 2025 they ask organisations with an AI chatbot in customer service to:
- offer an option for human contact alongside the chatbot
- stay in control of the information the chatbot provides
- make sure users know they are communicating with an AI system
On 14 April 2026, the ACM and AP asked for input on a guideline on chatbots in customer service. They expected to adopt the final guideline in autumn 2026. Check acm.nl to see whether it has been published.
So build in a handover to a team member, with the chat history attached, so the customer does not have to type the question again. Outside office hours, the agent says when a person will respond. How to set up such an agent is covered in the article on an AI agent for customer service.
What applies to AI images, deepfakes and AI text?
Marking is a provider obligation. Generators of images, audio, video and text mark their output in a machine-readable format, for example with a watermark or metadata. According to the Commission, systems placed on the market before 2 August 2026 must comply from 2 December 2026. On 10 June 2026 the Commission published a code of practice on marking and labelling AI-generated content, which can be signed voluntarily.
You disclose deepfakes yourself. A deepfake is AI-generated image, audio or video content that resembles existing persons, places or events and would falsely appear authentic. If you have an AI version of yourself record a video, disclose that visibly or audibly, at the latest at first exposure. The tool's machine-readable watermark does not count for this.
You disclose AI text in one case. The obligation applies to text you publish to inform the public on matters of public interest, such as politics, public health or consumer safety. If a human has reviewed the text substantively and someone holds editorial responsibility, the obligation falls away. A spell check does not count.
What happens if you do not comply with Article 50?
Under Article 99, a breach of Article 50 carries a fine of up to 15 million euros or 3 percent of worldwide annual turnover. For SMEs, the lower of the two amounts applies.
Enforcement lies with national market surveillance authorities. In the Netherlands, no authority has been formally designated to impose AI Act fines as of November 2026. The Dutch government held a consultation on the AI Act implementation bill from 20 April to 1 June 2026 and, according to the digital affairs planning letter, plans to submit it to the House of Representatives in the second quarter of 2027. The obligation itself has applied since 2 August 2026, and the GDPR and consumer law apply alongside it.
Frequently asked questions
Do I have to disclose that my customer service uses AI? Yes. Since 2 August 2026, Article 50(1) of the AI Act requires that people know they are communicating with an AI system, unless this is obvious. Put the disclosure in the first message of your chat or WhatsApp and in the opening sentence of a voice agent.
Who is responsible for the AI disclosure: me or my software vendor? Article 50(1) places the obligation on the provider, usually the vendor of the tool. If you have your own AI agent built under your own name, you can be the provider yourself. Either way, check yourself that the disclosure is switched on.
Must a customer be able to reach a team member? Article 50 does not require it. In their opinion of October 2025, the ACM and the AP do state that organisations should offer human contact alongside an AI chatbot.
When do AI images need a watermark? Providers have marked AI output in a machine-readable format since 2 August 2026. For systems already on the market before that date, this applies from 2 December 2026. As a deployer, you also disclose deepfakes visibly or audibly.
How high is the fine for breaching Article 50? Up to 15 million euros or 3 percent of worldwide annual turnover; for SMEs the lower amount applies. In the Netherlands, no authority has been formally designated to impose these fines as of November 2026.
AI disclosure and handover built in by default
With an AI agent implementation, from €1,495 one-off, the AI disclosure in the first message and the handover to a team member are built in by default for each channel. Want to know first where your team stands on AI knowledge? The free AI literacy test has 55 questions and gives a certificate from 70 percent. Legally you do not need that certificate: Article 4 asks for measures that support AI literacy, and according to the European Commission no certificate is required.
