From 2 August 2026 a website chatbot must say it is a bot
If a bot answers the chat on your website, from 2 August 2026 it has to tell the visitor it is not a person, no later than its first message. Article 50 of the EU regulation on artificial intelligence requires it. The fix is free and takes five minutes: write into the bot's greeting what it is and how to reach a real person.
A chat bot is like a shop where a tape recorder stands in for the assistant. The recording answers quickly and always politely, until the customer asks something that isn't on it. Now the tape recorder has a duty: to say that it is a tape recorder.
A chat widget sits on every tenth site
In September 2026 I went through 27,591 working Latvian company websites and found a chat widget on 2,923 of them (10.6%). That's the little window in the corner of the page through which a visitor writes to the company. Who answers it, a person or a program, my check can't see: it only records that the window is there.
On some of them a person answers in working hours; on others a bot answers round the clock. The visitor can't tell the difference, and that's exactly what the law is about.
The law is about the first message
Article 50(1) of Regulation (EU) 2024/1689 on artificial intelligence (known as the AI Act) requires that an AI system which talks directly to people is designed so that the person understands they are talking to an AI. There is one exception: when that is obvious to a reasonably attentive person, in the circumstances. Article 50(5) pins down the moment: the information must be given clearly and distinguishably, at the latest at the first interaction. Under Article 113, Article 50 applies from 2 August 2026.
In summer 2026 the regulation was amended (Regulation (EU) 2026/1744 of 8 July 2026): the deadlines for high-risk systems moved to 2 December 2027 and 2 August 2028, and the date for Article 50 was left alone. The only postponement concerns paragraph 2, the machine-readable marking of generated text and images: providers of systems placed on the market before 2 August 2026 got until 2 December 2026. That is a duty of whoever builds the bot, not of whoever has one on their site. All of this is law in force, not plans.
Formally, Article 50(1) is addressed to the provider of the system, the service that built the bot. But the visitor sees your site and your name above the window. My opinion: if the widget doesn't introduce itself, the customer will hold you to account, and you will then hold the provider to account. This is not legal advice; the text of the article is linked below, and on 15 October 2026 the Datu valsts inspekcija (DVI, the Latvian Data State Inspectorate) opened a section on its website about supervision of the AI regulation and promises to publish guidance.
What to do
- Put in the bot's first message what it is. In the chat service's admin panel, find the greeting field, the first message the visitor sees before they type anything. What that field is called depends on the service. The text: "Hello! I'm an automated assistant. I answer straight away, but I'm not a person. If you need a person, type 'operator' or call +371 XXXX XXXX, weekdays 9 to 6." In three languages, if your site is in three.
- Leave a way to a person in the chat window. A phone number and email in the greeting or in the widget's header, and the same number as on the site, as text and as a call link. A tape recorder is honest when a note beside it says where to find the live assistant and when they're in.
- Check that the bot doesn't collect more than your form does. Ask the chat service what it stores (the conversation, the visitor's address, anything else) and for how long. Add the chat to your privacy policy next to the form: a visitor who leaves you a phone number has the right to know who gets it (Article 13(1) of the GDPR). If the bot asks for a phone number and email "to follow up", let it ask for exactly what the form asks for.
- Read every text a neural network wrote, and put the company's name under it. Article 50(4) requires disclosing that a text was generated by AI if it is published to inform the public on matters of public interest. I won't try to say whether your article about choosing a boiler falls under that description. But the exception in the same paragraph is clear: there is no duty if the text has been through human review or editorial control and a specific person, natural or legal, is responsible for publishing it. Make a rule: no text goes on the site until a person has read it, and the page says who is responsible for it. That is editorial control.
- Don't put up generated photos of "the team". The same paragraph 4 requires disclosing deep fakes. Under Article 3(60) these are generated images that appreciably resemble existing persons, places or events; a photo of a "team" that does not exist may well fall outside that definition. But even without the law: a customer who turns up after seeing such photos will meet real people and draw a conclusion about everything else. One honest phone photo is better.
What it buys you
A visitor who writes to the chat in the evening gets an answer and knows who it's from. In the morning they phone you and don't expect you to remember yesterday's conversation with "the manager". How this affects orders, I haven't measured.
How I counted
September 2026, 27,591 Latvian company websites that responded during the check. The "chat widget" mark means the program found a chat window in the page code; who answers in it and whether the bot introduces itself, I did not check. The status of the regulation and its amendments is taken from the texts on EUR-Lex. More: how I checked 27,591 websites.
Sources
- My own check, September 2026: a chat widget on 2,923 of 27,591 sites.
- Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, Article 50: paragraph 1 (a system talking to people says it is AI), paragraph 4 (disclosure of generated text and images, exception for editorial control), paragraph 5 (at the latest at the first interaction).
- Regulation (EU) 2024/1689, Article 113: applies from 2 August 2026; the text of the regulation on EUR-Lex.
- Regulation (EU) 2026/1744 of 8 July 2026: new deadlines for high-risk systems; the new paragraph 4 of Article 111: providers of generative systems placed on the market before 2 August 2026 comply with Article 50(2) by 2 December 2026.
- GDPR (Regulation 2016/679; in Latvia, VDAR), Article 13(1).
- DVI: Informācija par MI aktu un tā uzraudzību vienuviet (15 October 2026): the section on supervision of the AI regulation on the inspectorate's website.
If you would like me to write to your chat one evening and tell you who answered, how they introduced themselves and what they asked back, get in touch.
This article is for information only and is not legal advice. For your own situation, ask a lawyer or your data protection authority.