Do you have to tell customers they are talking to AI? EU AI Act Article 50 and Southeast Asia

For systems within the EU AI Act's scope, providers generally must design chatbots so users are told they are dealing with AI, unless that is obvious. Article 50(1) has applied since 2 August 2026. The Southeast Asian documents reviewed give guidance and do not by themselves establish a legal duty in every context. This is not legal advice.

This article reflects sources checked on 22 October 2026. A lawyer qualified in your jurisdiction should review your own set-up.

What does Article 50 of the AI Act require?

Article 50 of Regulation (EU) 2024/1689 (opens in a new tab)^[1] sets four transparency duties. The first is the one about talking to a chatbot. It says AI systems intended to interact directly with people must be designed so that people are "informed that they are interacting with an AI system, unless this is obvious". Obviousness is judged from the viewpoint of a reasonably well-informed, observant and circumspect person. The marking duty in paragraph 2 can also reach text that a generative system produces, which is why the dates below matter.

The Commission's guidelines on Article 50 (opens in a new tab)^[3], dated 20 July 2026, say the exception should be interpreted restrictively. They list AI chatbots in helpdesks, where users may perceive the replies as human-generated, among the examples where the exception does not apply. The guidelines call themselves non-binding; only the Court of Justice can give an authoritative reading.

Paragraph 5 sets the timing. Information must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure", and must meet applicable accessibility requirements. The guidelines add that a disclosure found only in terms and conditions is insufficient on its own: it may complement an in-context notice but not replace it. They call a single prominent notice "likely to suffice in most instances", with periodic reminders likely to be needed in riskier contexts such as financial, insurance, legal and health advice.

Are you the provider or the deployer?

A provider, under Article 3(3), develops an AI system or has one developed, and places it on the market or puts it into service under its own name or trademark. A deployer, under Article 3(4), uses an AI system under its authority, other than for personal non-professional activity.

The guidelines give examples. A company that builds a chatbot in-house and puts it into service in the EU under its own name is a provider. One organisation can be both. A company that modifies an existing system, for instance with new training data, and puts the result into service under its own name becomes the provider of the new system.

A business that configures a chatbot on someone else's platform sits between these cases, and the answer turns on facts such as who developed the system and whose name it carries. Who counts as provider or deployer is a legal question this article does not answer for any vendor, Plandora included. Put two questions to your vendor in writing: who supplies the notice shown at the first interaction (Article 50(1)), and who handles machine-readable marking of generated outputs (Article 50(2)).

Does AI-generated content outside the chat need a label?

Paragraph 4 falls on deployers, so on the business that publishes the content. It covers two cases.

The first is a deep fake under Article 3(60): AI-generated or manipulated image, audio or video that looks or sounds like real people, objects, places, entities or events, and that a person would wrongly take to be authentic or truthful. The guidelines give an AI-generated recording that copies the voices of a podcast's regular presenters, and a realistic synthetic avatar of a company CEO, as examples. They read "persons" to include a real person's image and voice, and they judge whether content falsely appears authentic from the whole context. A trained copy of an expert's own voice could therefore fall in scope, depending on how it is presented to visitors.

The second is AI-generated text published to inform the public on matters of public interest, such as a lifestyle-website article on diets. It is exempt where the text "has undergone a process of human review or editorial control" and someone holds editorial responsibility. The guidelines say text is not published when access is restricted to specific individuals, and list private, interpersonal correspondence for professional purposes and a chatbot news summary available only to the user who prompted it as examples. That points to a private one-to-one chat reply raising an Article 50(1) question, not an Article 50(4) one.

When does it apply, and what did the Digital Omnibus change?

Article 113 of the Act makes it apply from 2 August 2026. Regulation (EU) 2026/1744 (opens in a new tab)^[2], the Digital Omnibus on AI, is dated 8 July 2026 and was published in the Official Journal on 24 July 2026. It enters into force on the third day after publication, which makes 27 July 2026. It adds a paragraph 4 to Article 111: providers of generative systems that were on the market before 2 August 2026 "shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026".

The delay covers marking only. The guidelines say a system that is partly interactive and partly generative may use the transition for marking, but "compliance with the disclosure obligation for AI systems directly interacting with natural persons must be ensured as of 2 August 2026".

Commission pages word the delay differently. The Article 50 FAQ (opens in a new tab)^[4], which still showed a last update of 24 July 2026 when read on 22 October 2026, says a limited grace period "is envisaged". The quick-facts page (opens in a new tab)^[5], last updated on 29 July 2026, says Article 50(2) was "amended by AI Omnibus". The Official Journal text settles it: the 2 December 2026 date is in the published Regulation.

The Commission's Code of Practice on Transparency of AI-generated Content (opens in a new tab)^[6], whose final version the page dates to 10 June 2026, covers marking by providers and labelling by deployers. Signing is voluntary, and the Commission's page says the legal duties apply either way.

Does it apply to a business outside the EU?

It can. Article 2(1)(c) covers providers and deployers established in a third country "where the output produced by the AI system is used in the Union". Scope depends on the role. Under Article 2(1)(a), a provider that places a system on the EU market or puts it into service there is covered wherever it is established. The guidelines add two limits for third-country operators. A deployer is in scope where it foresees distribution and use of the output in the EU, for instance by directing or authorising it. For a provider whose system is not placed on the EU market, incidental, unforeseeable or unauthorised downstream use should not alone trigger the duties.

The guidelines do not apply this to chats, so what follows is a reading, not their text. A Southeast Asian expert who invites EU customers into a chat looks closer to foreseeable use. One whose site is merely reachable from Europe looks closer to incidental use. Which side a business falls on depends on its role and the facts.

What are the penalties?

Under Article 99(4)(g) of the Regulation (opens in a new tab)^[1], breaching the Article 50 duties can bring administrative fines of up to EUR 15 million or, for an undertaking, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs, including start-ups, Article 99(6) makes each fine up to the percentage or the amount, whichever is lower, and the Digital Omnibus adds the same rule for small mid-cap companies. These are ceilings. Member States set the penalty rules, and the Commission's FAQ (opens in a new tab)^[4] says compliance will mainly be enforced by national market surveillance authorities.

What do Singapore, Malaysia and Indonesia ask for?

In the documents reviewed, none is a statute with a general duty to tell users they are talking to AI. Other laws, such as consumer-protection and misrepresentation rules, may still matter and were outside this review.

IMDA's "declare upfront" wording sits in its framework for agentic AI, so the first question is whether a Digital Mind counts as an agent. The framework says there is "no consensus on what defines an AI agent" and describes agents as usually having "some degree of independent planning, decision-making, and action-taking" over multiple steps. Its examples of agents that users interact with include customer service and sales agents. A Digital Mind that only answers from approved knowledge sits further from that description than one set up with multi-step Flows, Specialists or Automations, so do not assume the wording is irrelevant to your set-up. This article does not settle where any given Digital Mind falls.

Singapore's PDPC says its guidelines are "advisory in nature, are not legally binding on the Commission or on any other party". The Personal Data Protection Act they explain is binding, and its notification duty matters for any chat that collects names, emails or phone numbers. The companion article on data protection for AI chat in Southeast Asia covers that side.

Malaysia's guidelines list "full disclosure about the facts that an AI system is being used in decision making" as the first of five transparency elements, and say the principle is mainly applicable where AI is part of a decision-making process. Appendix 8, an example code of ethics, says "Transparency shall also mean disclosing to users that they are interacting with an AI." That is example wording for organisations to adopt, in a document that calls itself voluntary.

Indonesia's circular letter is framed as ethical guidance (pedoman etika), and its text does not say users must be told they are talking to AI. Whether a circular letter binds anyone is a question for Indonesian counsel. Komdigi's own page of 12 February 2026 (opens in a new tab)^[13] describes Presidential Regulations on a national AI roadmap and on AI ethics as drafts, and it could not be confirmed from an official source that either has since been signed. No Malaysian AI statute was found in force either.

What does good disclosure look like in a Digital Mind chat?

What follows is good practice drawn from the Commission and IMDA documents, not compliance advice. A Digital Mind is not the human expert and should not be presented as one, so plain labelling fits.

1. Say it before the visitor types. Open with one plain sentence saying this is an AI assistant for your business. The Commission's guidelines list "Unclear or ambiguous signals (e.g. generic references to “assistant”)" as insufficient on their own.
2. Keep it visible. The same guidelines pair first-turn greetings with "persistent badges or labels", and IMDA's chatbot guidelines suggest keeping the info card reachable through a persistent point such as an information icon. Plandora's documentation does not describe a setting for a badge or icon, so check what visitors actually see where they chat and do not assume one is there. The documented settings you control are the Short bio under Mind › Profile and Playbook instructions for how the Mind replies. Use both to state that this is an AI assistant.
3. Show the way to a person. In Plandora, Playbook Boundaries can set when a question needs a person, Flow steps can tag the person or team who handles a handoff, and the Inbox lets them take over and later release the conversation back to the Mind. When a person takes over, tell the visitor, so they know who is writing. Human handoff covers the triggers.
4. Label a trained voice. Plandora's Voice lets an owner train their own voice so visitors can hear replies spoken aloud. As documented, a visitor starts playback by selecting "Hear me speak" on a text reply, so the audio follows a reply rather than opening with a spoken greeting. The Commission's example of a spoken opening, "This is an AI-powered assistant", is written for voice-based and telephony contexts, which the Voice documentation does not describe. If you use Voice, say in your Short bio and in the Mind's first reply that the voice is an AI voice trained on yours.
5. Repeat the reminder where stakes are higher. For health, money or legal questions, the guidelines say periodic reminders are likely to be necessary, and Plandora's own guidance is to keep a person involved in consequential decisions. Why AI replies go wrong explains where errors come from.
6. Test what a stranger sees. After any Playbook edit, wait for the Playbook to show Live, because blocked changes stay as a draft. Then open a direct chat link, such as the Quick Chat link, in a private window without visiting the profile first, and check that the AI notice appears where you intended. Test voice playback and a human takeover in the Inbox as separate checks. For each channel you use, record the notice a visitor actually saw.

Nothing here is legal advice. Before launch, have a lawyer qualified in your jurisdiction review your set-up.