AI Act
Do You Need to Tell a Client They're Talking to a Robot?
Yes. Since 2 August 2026, Article 50 of the EU AI Act requires that a person be clearly told they are interacting with artificial intelligence, no later than the time of the first interaction. This applies to chatbots on websites, voice agents on the phone, and AI-generated content. The penalty: up to EUR 15 million or 3% of annual turnover.
What exactly does Article 50 require?
The requirement is short: if a system is intended to interact directly with a person, that person needs to know they are talking to artificial intelligence. The information has to be given in a "clear and distinguishable manner" and no later than the time of the first interaction, not at the end of the conversation, and not buried in the seventh paragraph of the terms of use.
- Website chatbot — disclosure before the first reply, or in the window's header, not just in the privacy policy.
- Phone voice agent — in the first sentence, before it starts asking questions.
- AI-generated text, image, audio, or video — must be marked as artificially created.
- Emotion recognition or biometric categorisation — the person needs to be informed that such a system is operating.
The exception applies only when this is obvious to a person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use. In practice this is read narrowly: an ordinary customer-support bot that writes in natural language and is indistinguishable from a human doesn't get to hide behind this exception.
Which obligations apply from which date?
In July 2026 the EU adopted the "Digital Omnibus", Regulation (EU) 2026/1744, published on 24 July and entering into force on 27 July. It pushed back some deadlines, but not Article 50: the transparency obligations started on schedule, on 2 August 2026.
| Obligation | Status | Date |
|---|---|---|
| Prohibited practices (Article 5) | In force | February 2025 |
| AI literacy for staff (Article 4) | In force, wording softened by the Omnibus | 2 February 2025 |
| General-purpose AI model (GPAI) obligations | In force | August 2025 |
| Transparency: chatbots, voice agents, AI content (Article 50) | In force | 2 August 2026 |
| Machine-readable marking for systems already on the market before 2 August 2026 | Transition period | 2 December 2026 |
| High-risk systems, Annex III (standalone) | Postponed by the Omnibus | 2 December 2027 |
| High-risk systems, Annex I (embedded in products) | Postponed by the Omnibus | 2 August 2028 |
How large is the penalty?
For breaching the transparency obligations, the AI Act provides for a fine of up to EUR 15 million or up to 3% of the company's total worldwide annual turnover, whichever is higher. For a small company that in practice means a percentage of turnover, not millions, but the fine isn't the only risk: a client who discovers they were talking to a machine without being told usually doesn't file a complaint, they just leave.
Who oversees AI Act enforcement in Latvia?
Latvia doesn't have a single AI supervisory authority. The model is split: the Ministry of Economics is the notifying authority, the Latvian National Accreditation Bureau handles accreditation, the Ombudsman covers fundamental-rights protection, and market surveillance is divided among roughly thirteen authorities depending on the sector, including the Consumer Rights Protection Centre and the Data State Inspectorate.
In practice this means two things. First, the answer to "who do I call" depends on where the system is used. Second, as of this article's date, no separate public guidance specifically about chatbots has been published in Latvian, so a company has to work from the regulation's own text and the Commission's guidance.
How do you comply with this in a single day?
Technically this isn't much work. The hard part is wording it so it doesn't sound like a legal disclaimer.
- Chatbot — the first message at the top of the window: "This is the company's AI assistant. If you need a person, type 'human'." It satisfies the law and gives the client a way out at the same time.
- Voice agent — the first sentence after the greeting, before any questions. Not after the third minute.
- AI-generated images and video — a visible marking on publication and a machine-readable marking in the file; for systems that were on the market before 2 August 2026, the machine-readable part has until 2 December 2026.
- Human-reviewed text — under the Commission's guidance of 20 July 2026, a machine translation of human-written text is excluded from the marking requirement; fully generated content is not.
- Document what was done. If anyone asks, there needs to be something to show: a screenshot, a date, who was responsible.
What does this change in a conversation with a client?
Honestly, trust drops from the first second. In AnswerConnect's April 2026 survey (around 6,000 people), 85% said they prefer talking to a human, and 31% hang up once they realize they're talking to a bot. A year earlier those figures were 83% and 29%, so the trend is moving up, not down.
So the disclosure isn't just a box to tick for compliance. It changes the script: the bot needs to be brief, offer a quick way to reach a human, and only promise what it can actually deliver. A system that reveals in the first line that it's AI and immediately offers to switch to a human annoys people less than one that pretends to be human for five minutes.
How we did this ourselves
This website runs our own AI assistant, and it tells you in the first line that it's a machine and offers to hand off to a human. That's not a last-minute patch: when we build a website chatbot or a voice and email AI agent for a client, the disclosure is part of the work, not an add-on service.
If you already have a bot running and aren't sure whether it meets the requirements, the first AI consultation is free: we review where the system interacts with people and tell you what needs fixing. It's not a legal opinion, just a practical checklist.
Sources
- AI Act Article 50: transparency obligations (artificialintelligenceact.eu) (05.09.2026.)
- European Commission AI Act Service Desk on Article 50 (05.09.2026.)
- AI Act Article 4: AI literacy (05.09.2026.)
- Digital Omnibus and postponed deadlines (Mayer Brown, July 2026) (05.09.2026.)
- National implementation plans: supervisory authorities in Latvia (05.09.2026.)
- VARAM on Latvia's AI Act implementation plan (05.09.2026.)
- AnswerConnect survey on attitudes toward bots (April 2026) (05.09.2026.)
Frequently asked questions
From what date must you disclose that AI is responding?
Since 2 August 2026. That's when the transparency obligations of Article 50 of the EU AI Act took effect. For systems that were on the market before that date, machine-readable marking of AI content has a transition period until 2 December 2026.
Does this also apply to small companies?
Yes. Article 50 sets no exemption based on company size. If an AI system talks to a client on your website or by phone, the disclosure is needed whether your company has three employees or three hundred.
Where exactly does it need to say it's a robot?
Wherever a person sees it before the conversation starts: in the chatbot window's first message or header, and for a voice agent, in the first sentence. A line in the privacy policy or terms of use doesn't satisfy the requirement, because that isn't a "clear and distinguishable manner" at the time of the first interaction.
How large is the penalty for not disclosing it?
Up to EUR 15 million or up to 3% of worldwide annual turnover, whichever figure is higher.
Does AI-generated content need to be marked?
Yes, artificially generated or manipulated text, image, audio, and video must be marked as such. Under the European Commission's guidance of 20 July 2026, a machine translation of human-written text is excluded from this requirement, but fully generated content is not.
Who checks AI Act compliance in Latvia?
There's no single responsible authority. The Ministry of Economics is the notifying authority, the Latvian National Accreditation Bureau handles accreditation, the Ombudsman oversees fundamental rights, and market surveillance is divided among roughly thirteen sector authorities, including the Consumer Rights Protection Centre and the Data State Inspectorate.
Is Article 4 on AI literacy still in force?
Yes, it has been in force since 2 February 2025. The Digital Omnibus (Regulation (EU) 2026/1744) softened the wording: it now requires "taking measures" rather than ensuring, to the best of its ability, a sufficient level of AI literacy, but it did not repeal the article and did not move the date.
Is this article legal advice?
No. This is a practical explanation with references to the regulation's text and public sources. For your specific situation, especially if the system touches health data, finances, or children, you need a lawyer.
Let's check your bot together
If AI is already running on your website or phone line, in 90 minutes we'll walk through where the system meets people and tell you what needs fixing to comply with Article 50. The conversation is free, no strings attached.