Documents
AI Transparency
Public page of the Chattopus service, provided by ERBA INTERACTIVE, on transparency about artificial intelligence (AI Act, Article 50). Version 1, updated on 19 November 2026.
In short
- From 2 August 2026, whoever provides a chat based on artificial intelligence must let the people who use it know that they are talking to an AI, at the latest at the first interaction (Article 50(1) of Regulation (EU) 2024/1689, the “AI Act”).
- Chattopus shows this disclosure every time the artificial intelligence answers on your website: the “AI” mark next to the assistant's name, at the top of the chat, with the full sentence for people who use a screen reader or hover over the mark; with whichever AI provider you have connected, and even without our mark. It cannot be switched off.
- Automatic replies, taken from your FAQs or written by you for the quick-reply buttons, are not written by the AI: you can recognise them by the label “Automatic reply”. When only the people on your team answer on the website, the disclosure does not appear, because there is no AI.
- For each conversation we record when the disclosure was shown, and we mark the assistant's replies as AI-generated.
- You, as the customer, must not remove or circumvent it. This page explains what we do and what we ask you to do.
The rules involved
| Rule | What it says, in short | Applies from |
|---|---|---|
| AI Act, Article 50(1) | The provider of an AI system that interacts with people must design it so that people know they are interacting with an AI, unless this is obvious to a reasonably well-informed person | 2 August 2026, with no extension |
| AI Act, Article 50(2) | The provider of a system that generates text, images, audio or video must mark the generated content in a machine-readable format, so that it can be recognised as artificially generated | 2 August 2026 for new systems; 2 December 2026 for those already on the market before 2 August 2026 |
| AI Act, Article 50(5) | The information must be given in a clear and distinguishable manner, at the latest at the first interaction, and must meet the accessibility requirements | 2 August 2026 |
| AI Act, Article 3(3) and (4) | A “provider” is whoever places an AI system on the market or puts it into service under its own name or trademark; a “deployer” is whoever uses it under its own authority in the course of a professional activity | — |
| AI Act, Article 99(4) and (6) | Fines of up to €15 million or 3% of total worldwide annual turnover; for SMEs the lower of the two amounts applies | 2 August 2026 |
| AI Act, Article 4 | Providers and deployers take measures to ensure a sufficient level of AI literacy of their staff | 2 February 2025 |
| Italian Law 132/2025, Article 4 | Access by children under 14 to AI technologies requires the consent of whoever exercises parental responsibility | 10 October 2025 |
Who does what
- ERBA INTERACTIVE is the provider of Chattopus: it designs the chat so that, when the AI answers, the disclosure and the marking are always there.
- You, the customer, use the chat on your website for your business: you are the deployer. Your main task is not to obstruct the transparency measures and to use the chat as intended.
- If you are an agency and resell the chat to your clients under your own brand, for that offer you are very probably a provider too, within the meaning of Article 3(3). The details are in the Agency and Reseller Addendum.
- The provider of the model you have connected with your key (Anthropic for Claude, OpenAI for ChatGPT, Google for Gemini, xAI for Grok, Mistral AI, DeepSeek, Groq, OpenRouter or another service compatible with the OpenAI APIs) is the provider of the general-purpose AI model, with its own obligations under the AI Act; ERBA INTERACTIVE remains the provider of the chat that uses it, and the transparency measures described here apply whatever model you choose.
What Chattopus does for Article 50(1)
- The disclosure is always there when the AI answers. When the artificial intelligence answers on the website, the chat receives from the server, together with the website's configuration, a disclosure sentence that is never empty, in the chat's language. The chat currently speaks English, Italian, French, Spanish and German: of the languages the visitor's browser lists, it uses the first one it knows, otherwise the website's default language, and always the website's language if you have switched on “Chat always in Italian” (or in the language you chose for the website) in the website's settings. The installation code can ask for a language (the
data-localeattribute, which the WordPress plugin adds only if a developer asks for it): the chat uses it if it knows it, otherwise it uses the website's language. The disclosure, the chat's texts and the language saved with the conversation are always the same language. - It is not one of the texts you can customise. From the dashboard you can change the welcome messages, the quick-reply buttons and their replies, the button texts, the messages for out of hours and for when the assistant is not available, the placeholder of the message box, the look of the chat (colours, backgrounds, fonts, images, free elements, order of items, animations) and the styles (CSS), but not the disclosure: it is the “AI” mark next to the assistant's name, in the chat's header, and free elements do not cover it.
- It stays even without our mark. On the plans without the “Powered by Chattopus” mark, and with an agency's brand, our mark disappears, not the disclosure.
- At the first interaction. The “AI” mark is in the chat's header as soon as the window opens, before the greeting and before the visitor writes, and stays visible while the AI answers (it does not scroll away with the messages). The default welcome message also says that an artificial intelligence is answering, but the Customer can rewrite it.
- It is recorded. With each message the chat tells the server that the disclosure was shown, and the server saves for each conversation the moment this happened. You find it in the data export (Organization page, “Download data”).
- AI and people are kept apart. The assistant's replies and those of the people on your team are shown as different types of message; messages from people show the name of whoever writes them.
- A person can always be reached. The visitor can ask for one at any time: by writing so in the chat, with “Talk to a person” when the Customer shows it at the opening or when the assistant offers it, or with “Would you rather talk to a person?” inside “Your contact details”, the button at the top. The request is sent at once and does not require the visitor to leave anything: neither an email address nor a phone number. Afterwards, the chat offers the visitor the option of leaving their email address to receive the reply there too. Outside the team's hours, if you have set them, the request still arrives: the person who makes it reads your out-of-hours message, and the request waits for your team in the Inbox, without a notification (chats do not arrive by email); the team replies in the chat when it is back.
- Accessible. The disclosure is real text in the page: the “AI” mark is linked to the chat window as its description (
aria-describedby), so screen readers read the full sentence when the chat opens. - When the AI does not answer. If the website answers with automatic replies, that is with the text of one of your FAQs chosen by matching the words of the message, that message is labelled “Automatic reply” and does not carry the AI marking, because it is not generated by an AI. The same applies to the reply you wrote for a quick-reply button: the chat gives it when the visitor taps that button, even if the AI answers on the website. If the website answers only with the people on your team (for example on the Low plan), the chat says that the team will reply and does not show the disclosure.
What Chattopus does for Article 50(2)
- Every reply generated by the assistant is saved with an explicit indication that it was generated by AI. The indication appears in the dashboard and in the JSON data export, so it stays even when conversations leave the platform.
- In the chat's public interface, the assistant's replies are of a different type from those of people.
- In the chat, every reply from the assistant carries the machine-readable attribute
data-ai-generated="true", and the “AI” mark the attributedata-ai-disclosure="true".
Note. It is not yet clear whether Article 50(2) applies to the text replies of a chat, which the visitor reads already knowing that they are talking to an AI. We apply the marking anyway, as a precaution.
What you must not remove or alter
You must not, either directly or through whoever manages your website:
- hide, cover, shrink until illegible or remove the “AI” mark of the disclosure, for example with the dashboard's custom styles (CSS), with colours that make it illegible, or with CSS or scripts added to the page;
- hide the “Automatic reply” label or pass off an automatic reply as a reply from a person, including by giving the assistant the name or photo of a team member;
- modify the chat or interfere with how it works in order to remove the disclosure or the marking;
- ask us to switch off the disclosure: we cannot do so;
- remove the “AI-generated” indication from exported conversations if you republish them or pass them on to third parties;
- present the assistant as a person: in its name, in the texts or in the contents of the knowledge base;
- make the assistant's replies appear to be written by a person, or the other way round.
A breach of these obligations is a breach of the Terms of Service (Article 12.4).
What we ask you to do
- The assistant's name. If you give it a person's name (for example “Lucy”), always pair it with “virtual assistant” (for example “Lucy, virtual assistant”).
- Knowledge base. Do not add instructions or texts that make the assistant say that it is a person.
- Privacy notice. Explain in the chat's privacy notice that the replies are generated by AI. Our notice template already does so.
- Handing over to a person. Organise your team so that visitors' requests get an answer; if you are often unreachable, set the team hours (on the plans that include them) so that the chat collects the requests out of hours.
- Training. Explain to the people who use the dashboard how the assistant works, that it can make mistakes and how to take over a conversation.
- Children. Do not use the chat on websites or services aimed mainly at children under 14.
- Excluded uses. Do not use the chat in the high-risk use cases of the AI Act (for example recruitment or creditworthiness assessment).
- Reports. If you see that the disclosure does not appear, or appears incorrectly, write to assistance@chattopus.com straight away.
The mark and the sentence
In the chat: AI next to the assistant's name (IA in Italian). The full sentence, which the mark carries as its description (screen readers read it, and it appears when you hover over the mark with the mouse), is the following, in the chat's language:
Italian
Stai parlando con un assistente virtuale basato su intelligenza artificiale: le risposte sono generate automaticamente. Puoi chiedere di parlare con una persona in qualsiasi momento.
English
You are chatting with an AI virtual assistant: replies are generated automatically. You can ask to talk to a person at any time.
Short version, for small spaces (for example under the assistant's name):
- IT: "Assistente virtuale con intelligenza artificiale · Puoi chiedere una persona in qualsiasi momento"
- EN: "AI-powered virtual assistant · You can ask for a person at any time"
French, Spanish and German. When the chat is in one of these languages, it shows the sentences below. In German the mark is KI, the German word for AI; in French and Spanish it is IA.
- FR, the full sentence: "Vous discutez avec un assistant virtuel basé sur l'intelligence artificielle : les réponses sont générées automatiquement. Vous pouvez demander à parler à une personne à tout moment."
- FR, the short version: "Assistant virtuel basé sur l'intelligence artificielle : vous pouvez demander à parler à une personne à tout moment."
- ES, the full sentence: "Estás hablando con un asistente virtual basado en inteligencia artificial: las respuestas se generan automáticamente. Puedes pedir hablar con una persona en cualquier momento."
- ES, the short version: "Asistente virtual basado en inteligencia artificial: puedes pedir hablar con una persona en cualquier momento."
- DE, the full sentence: "Sie chatten mit einem virtuellen Assistenten, der auf künstlicher Intelligenz basiert: Die Antworten werden automatisch erzeugt. Sie können jederzeit darum bitten, mit einer Person zu sprechen."
- DE, the short version: "Virtueller Assistent auf Basis künstlicher Intelligenz: Sie können jederzeit nach einer Person fragen."
When a person takes over the conversation (recommended system message):
- IT: "Ora ti risponde una persona del nostro team."
- EN: "A member of our team is now replying."
Frequently asked questions
My website already says that the chat is a bot. Can I remove the disclosure? No. When the AI answers on the website, the disclosure must appear in the chat itself, at the first interaction, in every case: it is the “AI” mark at the top, next to the name.
Can I change the text of the disclosure? Not from the dashboard. If you have a particular need (for example a different language), write to assistance@chattopus.com.
With white label, does the disclosure stay? Yes. The brand shown is the agency's, but the disclosure always stays.
Can the visitor talk to a person? Yes, always. Even when the artificial intelligence is not available (an invalid key, credit used up with your provider, a limit you have switched on for the website reached), the chat stays open with the automatic replies and with your team.
Do automatic replies, from the FAQs or from the quick-reply buttons, need the disclosure? No: they are not generated by an AI, but carry the text you wrote yourself. We label them “Automatic reply” anyway, so that the visitor knows that a person is not answering at that moment.
Does anything change if I use another AI provider, for example ChatGPT or Gemini? No. The disclosure and the marking are the same with whichever provider and model you connect.
And the dashboard's Help? The dashboard's Help (an automatic assistant), which explains how to use the service, talks to you and your team, not to visitors: we present it as an automatic assistant, and its replies may contain mistakes. The panel opens with “Hi! I am Help, the dashboard’s automatic assistant” and warns “I can make mistakes”. When the replies are written by an artificial intelligence, the line “You are chatting with an AI.” also appears under the panel's title; when Help answers only with parts of our guide, that line is not there.
This page supplements the Terms of Service (Articles 12.4 and 14). It is not legal advice on your specific situation.