If an AI assistant answers your customers on your website, in Messenger or over the phone, a new rule has applied to it since 2 August 2026: people must know they are not talking to a human. This is Article 50 of the EU Artificial Intelligence Act. The obligation is simple and the fix usually takes minutes, yet many chatbots still introduce themselves with a first name and a smiley rather than with what they are. Below: what the article actually says, who it applies to and how to check your own bot.
Paragraph 1 requires that AI systems intended to interact directly with people are designed so that the people concerned are informed they are interacting with an AI system. There is one exception: when this is obvious to a reasonably well-informed, observant person from the context. Paragraph 5 adds how: the information must be provided in a clear and distinguishable manner, at the latest at the time of the first interaction. A line in the privacy policy or a note hidden behind an "i" icon does not meet that. People have to see it when the conversation starts.
Legally, paragraph 1 is addressed to the provider — whoever developed the AI system and offers it under their own name. If you use an off-the-shelf chatbot service, the provider is the company behind that service. But if you built the bot yourself, or had one built for you on the ChatGPT, Claude or another model API and it runs under your name, you become the provider. In practice the difference is smaller than it looks: the bot speaks on your website under your brand, and a customer who feels misled will complain to you. So it is worth checking either way, and asking any vendor how they handle it.
The rule applies wherever AI itself generates replies to a person: a chat window on a website, automatic replies in Messenger, Instagram or WhatsApp, a voice assistant answering the phone, AI that answers emails without a human reviewing them. It does not apply when AI only prepares a draft that a person reads and sends in their own name — the customer is then dealing with a human. A simple button menu with pre-written answers is most likely not an AI system at all, since its behaviour is fully defined by people. If you are unsure which group your tool falls into, the safest choice is to treat it as AI — a disclosure costs nothing.
One sentence visible from the first message is enough: "You are chatting with an AI assistant" or "I am an AI assistant answering from our website's information". The bot can have a name, but then it must say clearly next to that name that it is AI, because a human name without explanation is exactly what creates the impression of a person. A voice bot must say it at the very start of the call, not after a few questions. When a real employee takes over, show that too — customers see when a human is answering and trust both more. The notice must be in the customer's language: if the bot replies in Latvian or English, the disclosure should too.
Article 99(4) of the AI Act sets fines for breaches of Article 50 of up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs, Article 99(6) reverses this: whichever is lower. National market surveillance authorities enforce it; in Lithuania that role is coordinated by the Communications Regulatory Authority (RRT). In reality a fine reflects the nature and duration of the breach and is not automatically the maximum. But when the fix is one sentence in a chat window, even a small risk is hard to justify.
Article 50 has other paragraphs, but they matter less to most businesses. Paragraph 2 requires makers of AI tools to mark generated text, images and audio in a machine-readable way; the Digital Omnibus that entered into force in July 2026 gave systems already on the market until 2 December 2026. Paragraph 4 concerns users: AI-generated deepfake images or video must be labelled, as must AI-generated text published to inform the public on matters of public interest, unless a human reviewed it and takes editorial responsibility. An ordinary company blog or product descriptions reviewed by staff generally fall outside this. Importantly, the Digital Omnibus did not postpone the chatbot disclosure duty — it has applied since 2 August.
Open your website in a private browser window and start a conversation as a customer. One: is it visible before or with the first reply that this is AI? Two: is it stated in words, not only with an icon or a colour? Three: if the bot has a human name, does it say next to it that it is an AI assistant? Four: ask in another language — does the disclosure switch language too? Five: check Messenger, WhatsApp and phone bots if you have them. If any answer is "no", it is fixed in the bot's settings or with one email to the vendor.
In our own AI assistant, SmartChat, the disclosure is built into the chat window itself so it cannot be forgotten or switched off: "You are chatting with an AI assistant" is always shown at the bottom of the chat window, and if the assistant has a name, the header says "AI assistant" right under it. The text follows the customer's language automatically — Lithuanian, English, Latvian and Estonian. You can see it live in the IGLU Smart online store at iglusmart.eu. If you use a different bot and are not sure it complies, get in touch — we will take a look for free.
Send us your website address — we will check whether your chatbot discloses properly in every language and channel, and tell you what to change. Free of charge.
If you are planning an AI assistant or other automation, see our business systems work.
Frequently asked
No. The Digital Omnibus that entered into force in July 2026 postponed the high-risk requirements and gave systems already on the market until 2 December 2026 only for machine-readable marking of AI content. The chatbot disclosure duty has applied since 2 August 2026.
No. Article 50(5) requires the information to be clear and distinguishable and given at the latest at the first interaction. People must see it in the chat window when the conversation starts, not find it in a separate document.
Yes, but it must say clearly next to the name that it is an AI assistant. A human name with no explanation creates the impression of a person, which is exactly what Article 50 is meant to prevent.
National market surveillance authorities. In Lithuania the role is coordinated by the Communications Regulatory Authority (RRT), which is also the single point of contact.
