EU AI Act Article 50: what it means for your chatbot

Under EU AI Act Article 50, a chatbot or voice agent that talks to people must tell them they are dealing with an AI, at the latest at the first interaction, unless that is obvious. For a customer service bot in the EU this has applied since 2 August 2026, and according to the European Commission that includes bots that were already live. This guide covers what Article 50 asks of chatbots and voice agents, who is responsible, the exceptions, the dates after the Digital Omnibus, the fines, how it sits next to GDPR, and a checklist for before launch.

Legal status checked on 3 November 2026 against Regulation (EU) 2024/1689 as amended, and the Commission's guidance. This is not legal advice: check your own case with a lawyer.

What Article 50 requires of a chatbot

Article 50 has five operative paragraphs. For a customer-facing chatbot or voice agent, one always applies and the others depend on what your agent does.

Article 50 comes on top of everything else in the Act. It does not replace the high-risk rules if your use case is high-risk (Article 50(6)), and a disclosure does not make a practice lawful that Article 5 prohibits (recital 137).

Who is responsible: provider or deployer

The AI disclosure duty in Article 50(1) sits with the provider: whoever develops the AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark (Article 3(3)). The deployer is whoever uses the system under its authority (Article 3(4)).

In practice:

• You built the chatbot yourself, for example on a model API, and run it on your website under your brand: the Commission's guidelines give an in-house chatbot put into service under the organisation's own name as an example of a provider. Article 50(1) is yours.
• You use a vendor's chatbot product as it is: the vendor is the provider and must build the disclosure in. You are the deployer. You still configure the greeting, the persona and the languages, so check that the disclosure survives your configuration.
• You take a vendor's system, modify it and put it into service under your own name: the guidelines say you can become the provider of the new system.

Where the line falls for white-label and heavily customised bots is not settled for every case. Write down which role you think you hold for each chatbot and voice agent, and why.

When is it "obvious" that people are talking to an AI?

Article 50(1) does not apply where it is obvious to a "reasonably well-informed, observant and circumspect" person that they are talking to an AI. The Commission's guidelines on Article 50 say this exception must be read restrictively and is limited to cases where "almost no doubt" is left.

• Not obvious, so disclose: the guidelines name AI chatbots in online platforms and helpdesks, where users get replies they may take for human ones. A public customer service bot falls here. Public audiences include children, older people and people with lower AI literacy, which lowers the bar further.
• Can be obvious: code assistants used only by developers, and an internal assistant for trained, AI-literate staff.
• Weighs against obvious: a human name and photo, a natural-sounding voice, and personalised conversation.

The general public knowing that chatbots exist does not make a specific interaction obvious, according to the guidelines.

How to disclose, and what does not count

Article 50 does not prescribe a technique. The guidelines describe what works and what does not:

• Works: a first-turn message such as "You are chatting with an AI system", a label near the input field, a persistent "AI" badge. For voice: a spoken statement at the start of the call, such as "This is an AI-powered assistant".
• Not enough on its own: a line in the terms and conditions or a linked page; metadata or watermarks the user cannot see; a vague word like "assistant"; a site-wide "services on this website use AI"; "this system uses LLMs" without saying the user is talking to an AI; an audio tone without words.
• Repeat it in longer conversations and in contexts where people are more easily misled: the guidelines name financial advice, insurance, legal help, health and complaints handling.
• Answer truthfully when asked. The guidelines say the system must disclose whenever someone asks about its nature, or seems confused about whether a human is replying.
• Mixed conversations: where AI replies and human replies are blended, the AI-generated parts need disclosure unless a human properly reviewed and sent them.
• AI agents that act (bookings, purchases, correspondence) should disclose both that they are AI and on whose behalf they act.

The disclosure also has to meet accessibility requirements (Article 50(5)): a screen reader must be able to read it, and a voice agent must say it, not only show it.

Voice agents: three extra points

1. Say it at the start of the call. A spoken disclosure in the opening sentence, repeated if the call is long or the agent's role changes.
2. Synthetic speech may need marking. Article 50(2) covers synthetic audio. If you are the provider of a system that generates speech, its output must be marked in a machine-readable way, as far as technically feasible. Ask your speech vendor what marking it applies. Systems already on the market before 2 August 2026 have until 2 December 2026 for this paragraph (Article 111(4)).
3. Inferring emotions from the caller's voice changes the picture. An emotion recognition system identifies or infers emotions on the basis of biometric data (Article 3(39)). If your voice agent does that, the deployer must tell callers (Article 50(3)), and emotion recognition is also listed as high-risk in Annex III, point 1(c). Using it on your own call centre staff is prohibited under Article 5(1)(f), except for medical or safety reasons.

Since when: the dates after the Digital Omnibus

The Digital Omnibus on AI (Regulation (EU) 2026/1744) amended the AI Act and has been in force since 27 July 2026 (Commission announcement). It moved the high-risk dates but did not move the chatbot disclosure duty.

The AI Act has no transition rule for Article 50(1): the only Article 50 transition in Article 111 is for machine-readable marking under 50(2). The Commission's guidelines (point 153) say every in-scope system must comply with Article 50 from 2 August 2026, "regardless of their date of placement on the market", and that partly interactive, partly generative systems must disclose from that date. That is the Commission's reading, not a court ruling: before the guidelines, at least one law firm (Hogan Lovells, 2024) argued that systems placed on the market before 2 August 2026 might fall outside Article 50. The safe course is to disclose on chatbots that were already live.

Is a customer service chatbot high-risk?

Usually not. Answering questions, tracking orders or changing an address is not on the Annex III list, so Article 50 and AI literacy are the AI Act duties that apply. It becomes a high-risk case when the system itself is used for a listed purpose, for example:

• evaluating creditworthiness or setting a credit score (point 5(b); fraud detection is excluded);
• risk assessment and pricing for life and health insurance (point 5(c));
• deciding on eligibility for public assistance, used by or for public authorities (point 5(a));
• evaluating and classifying emergency calls (point 5(d));
• screening job applicants or evaluating candidates (point 4(a));
• emotion recognition (point 1(c)).

Article 6(3) lets an Annex III system escape the high-risk label if it does not materially influence the decision, for example because it only performs a narrow procedural or preparatory task. That assessment must be documented before launch, and a system that profiles people is always high-risk. A bot that collects a loan application and passes it to a human is a different case from a bot that decides who gets the loan.

Fines under Article 99

Breaking Article 50 falls under Article 99(4)(g): up to EUR 15 million or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups it is whichever is lower (Article 99(6)); since the Omnibus, the same applies to small mid-caps for fines under paragraphs 4 and 5 (Article 99(6a)). These are maximums. Member States set the enforcement rules and designate the authorities. In the Netherlands, the government intends the Autoriteit Persoonsgegevens to supervise Article 50, but as far as we could verify, the Dutch implementing act was not yet in force on 3 November 2026.

Article 50 and GDPR

The AI Act does not affect the GDPR (Article 2(7) AI Act). A chatbot that handles customer conversations processes personal data, so both apply:

• Telling people it is an AI does not replace your privacy notice. The guidelines say Article 50(1) serves a different purpose than the GDPR information duties (Articles 13 and 14 GDPR). You need both.
• A bot that decides (refuses a refund, rejects an application) on its own, with legal or similarly significant effects, triggers Article 22 GDPR: in most cases a right to human intervention and to contest the decision.
• A bot that reads out another customer's address is a personal data breach. Article 33 GDPR then requires notifying the data protection authority within 72 hours, unless the breach is unlikely to result in a risk.
• Transcripts need a lawful basis, a retention period and access controls, like any other customer record. A data protection impact assessment (Article 35 GDPR) may be required, depending on the data and the scale.

What is still unclear

• Guidelines are not law. The Commission approved the content of its Article 50 guidelines on 20 July 2026 (C(2026) 5054). According to that decision they are formally adopted, and only then applicable, once all language versions are ready; as of 3 November 2026 we could not confirm that formal adoption had happened. They are non-binding, and only the Court of Justice can give an authoritative interpretation.
• Chatbots already live before 2 August 2026. The Commission says Article 50(1) applies to them from that date; an earlier law-firm reading disagreed. Not tested in court yet.
• Marking of chat text. How Article 50(2) applies to everyday chatbot replies is less settled than for images or audio. The Code of Practice on AI-generated content (final version published 10 June 2026, voluntary) is one recognised way to show compliance for marking and labelling.
• Provider or deployer for white-label and heavily customised bots, as above.
• National enforcement. Which authority acts, and how, depends on each Member State's implementing law, and several are still in progress.

What to do before launch

• Decide, per chatbot and voice agent, whether you are the provider or the deployer, and record why.
• Check whether any use case is on the Annex III list (credit, insurance pricing, recruitment, public benefits, emergency calls, emotion recognition). If it is, follow the high-risk route or document an Article 6(3) assessment.
• Make the first chat message say plainly that the customer is talking to an AI. For voice, say it in the opening sentence.
• Give the disclosure in every language the agent speaks, not only the default one.
• Do not rely on terms and conditions, the word "assistant", a site-wide notice or a sound alone.
• If the bot has a human name or photo, keep the AI disclosure explicit next to it.
• Make sure the agent never claims to be human, including when a customer asks directly or tries to make it role-play as a person.
• Repeat the disclosure in long conversations and in finance, insurance, complaints and health journeys.
• Mark which replies come from the AI when a human agent takes over or edits replies.
• Check accessibility: readable by screen readers, spoken on calls.
• Ask your model and speech vendors what machine-readable marking they apply to generated text and audio.
• If the agent infers emotions from voice or face, inform people and treat it as a high-risk case.
• Update the privacy notice for chat and call transcripts; set retention; check whether Article 22 or a DPIA applies.
• Have a breach procedure for the day the bot shows one customer's data to another.
• Train the staff who run and monitor the agent (AI literacy, Article 4).
• Keep evidence: transcripts or recordings showing the disclosure, per language and channel, with date and version.
• Re-check after every change to the model, prompts or knowledge base.

The wider launch checks (security, journeys, language) are in the AI chatbot go-live checklist.

Where testing helps

Article 50 is partly a design decision and partly a behaviour question. The greeting is easy to check once. What is harder to know is whether the agent still says it is an AI when a customer asks "am I talking to a real person?" in Portuguese, when someone tells it to play "Sarah from support", or after a prompt change three releases later. The same goes for GDPR: you only find out whether the bot reads out another customer's address by trying to make it.

AltaSec tests chatbots in chat today: security (prompt injection, jailbreaks, data-extraction attempts and pressure to break your rules, with every reply checked for leaked data), simulated customers running your real journeys, and language benchmarks in English, Romanian and European Portuguese, each checked by native speakers. "Always say you are an AI, never claim to be human" can be one of the rules we push against. For voice agents, we measure hearing, understanding and speaking on recorded calls today; voice security tests and simulations are next (voice agent testing).

Afonso maps findings to the EU AI Act and GDPR as input for your own assessment, never as a compliance verdict. The Passport is our verdict with evidence, not a certificate.

Frequently asked questions

Does Article 50 apply to a rule-based chatbot?

Not if it is not an AI system. The Commission's guidelines exclude simple automated responses that do not meet the AI system definition, such as rule-based quick replies. A bot built on a language model is an AI system.

Does a chatbot that launched before 2 August 2026 need a disclosure?

Very likely yes, and that is how we would plan. The AI Act gives no transition period for Article 50(1), and the Commission's guidelines say all in-scope systems must comply from 2 August 2026 regardless of when they were placed on the market. The only Article 50 transition, until 2 December 2026, covers machine-readable marking under 50(2). The guidelines are not binding and one law firm has argued for a narrower reading, so this is not fully settled until a court rules on it.

Can our chatbot have a human name?

The AI Act does not ban names. But the guidelines treat human-like representations that may mislead as insufficient on their own, and a human name and photo make it harder to argue the AI is obvious. Pair a persona with an explicit AI disclosure.

Do we have to disclose when our human agents use AI to draft replies?

Not under Article 50(1), as long as a human properly reviews and sends the replies: the guidelines treat that as indirect interaction. If AI replies go out without real review, the disclosure applies to those replies.

Does the disclosure have to be in the customer's language?

Article 50(5) requires the information to be clear and distinguishable to the people concerned. A disclosure in a language the customer does not read is unlikely to meet that, so give it in every language the agent serves.

Our chatbot vendor says the product is "AI Act compliant". Is that enough?

It covers what the vendor controls as provider. You still decide the greeting, persona, languages and journeys, and you hold the deployer duties and your GDPR duties. Check the disclosure in your own configuration, in every language and channel.

Does AltaSec certify AI Act compliance?

No. We map findings to the EU AI Act and GDPR as input for your own assessment. We do not issue certificates, seals or compliance verdicts, and no test proves an AI is right 100% of the time.