AI Act Article 50: telling people they are talking to a machine, robots included

Article 50 of the EU AI Act has applied since 2 August 2026. It requires AI systems that interact directly with people to be designed so that those people are told they are dealing with AI, unless that is genuinely obvious. The European Commission's guidelines, adopted on 20 July 2026, list humanoid robots and robotic companion pets among the systems covered. That includes talking robots at stands.

This article is general information, not legal advice. Contracts and specific deployments should be checked with a qualified adviser.

What Article 50 covers

The text of Article 50 on the AI Act Service Desk sets out four separate transparency duties:

1. Interactive AI systems (Article 50(1)). Providers must design systems that interact directly with people so that those people are informed they are interacting with AI. There are two exceptions: when this is obvious to a reasonably well-informed person, and certain systems authorised by law for criminal investigation.
2. Synthetic content (Article 50(2)). Providers of systems that generate audio, images, video or text must mark the output in a machine-readable way, so that it can be detected as artificially generated or manipulated.
3. Emotion recognition and biometric categorisation (Article 50(3)). Deployers, meaning the organisations using the system, must inform the people exposed to it.
4. Deepfakes and certain published text (Article 50(4)). Deployers must disclose deepfakes. They must also disclose AI-generated text published to inform the public on matters of public interest, unless a human has reviewed it and taken editorial responsibility.

Article 50(5) adds a common rule: the information must be clear and distinguishable, must be given at the latest at the first interaction or exposure, and must meet applicable accessibility requirements.

When the rules apply

According to the AI Act implementation timeline on the AI Act Service Desk, Article 50 started to apply on 2 August 2026. The timeline reflects the amendments made by the Digital Omnibus on AI, which are covered in more detail in our overview of AI Act deadlines for robots.

The Omnibus added one narrow transition. Under the amended Article 111 of the AI Act, providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to meet the marking duty in Article 50(2).

The Commission's guidelines stress that this grace period does not extend to the disclosure duty for interactive systems. A system that both talks and generates content may use the extra time only for marking its output. Telling people they are talking to AI was required from 2 August 2026.

Robots are named explicitly

The Commission published its guidelines on transparency obligations for providers and deployers of AI systems on 20 July 2026, after a draft and a public consultation in May and June. The guidelines are not binding law, but they show how the Commission expects the rules to be read.

They break Article 50(1) into four elements: an AI system, intended to interact, directly, with natural persons. Interaction can be spoken, visual or physical, and it does not have to be started by a human. Among the examples in scope, the guidelines list voice assistants, chatbots, (humanoid) robots and cobots, AI companions, robotic companion pets and AI avatars. Industrial robots working in a closed setting, not intended to interact with people, are excluded.

Why "it is obviously a robot" is a weak argument

It is tempting to assume that nobody could mistake a humanoid robot for a person, so no notice is needed. The guidelines make that harder to rely on, for several reasons:

• The exception should be read restrictively. It covers only cases where there is almost no doubt about the nature of the interaction for an average member of the audience.
• The audience includes everyone who can reasonably be expected to meet the system, not just the intended users. Systems operating in public spaces are mentioned specifically.
• If children, older people, people with disabilities or people with low AI literacy are likely to be present, the exception cannot be relied on to protect them.
• For physical systems, visible mechanics make the AI nature more obvious. Close imitation of the real thing makes it less obvious. For voice, a human-sounding voice counts against obviousness.

The examples where the exception does apply are narrow: coding assistants for professional developers, internal tools for trained staff, and characters in single-player video games. A robotic companion pet that closely resembles a real animal is given as an example where the duty still applies.

On a cautious reading of the guidelines, a robot that speaks with a natural voice to mixed crowds at a trade fair or shopping centre is unlikely to qualify for the exception. Disclosure is the safer default.

What a good disclosure looks like

The guidelines leave the technique open, but they give examples. For voice-based systems they suggest a spoken statement at the start of the interaction. In longer sessions, especially after interruptions, periodic reminders may be appropriate. Tones or sounds can help visually impaired people, but are not enough on their own. Text labels, persistent icons and combinations of text, audio and visuals are also described as good practice.

Some approaches are described as insufficient when used alone:

• a notice that appears only in terms and conditions, a manual or a URL;
• machine-readable markings that people cannot perceive at the point of interaction;
• vague signals such as calling the system an "assistant", or human-like presentation that may mislead;
• purely technical statements such as "this uses an LLM".

Timing also matters. The guidelines say the first interaction refers to every new person, not only the first visitor of the day. They also say the system should disclose its nature whenever it is asked whether it is AI, or when someone appears to be confused about it. For multilingual crowds, see our guide to speech, languages and noisy venues.

Who is responsible: provider or deployer?

Article 50(1) places the duty on the provider: the organisation that develops the AI system, or has it developed, and places it on the market or puts it into service under its own name. The guidelines add that a company which modifies an existing system and puts it into service under its own name becomes a provider of the new system. Who is the provider for a robot at an event, whether the robot maker, the software integrator or the rental company running a custom conversational layer, depends on the set-up. The guidelines do not settle every such case.

Deployers have their own duties under Article 50(3) and 50(4). An organiser who asks for a robot that reads visitors' emotions, or that speaks with a cloned voice of a real person, takes on obligations of their own. Some emotion recognition is banned outright in workplaces and education, as explained in our article on emotion-reading robots and the AI Act.

Teleoperated robots and synthetic content

Many event robots are partly or fully controlled by a human operator. In its scope analysis, the guidelines exclude mediated interaction, where a person relays or reviews AI output and remains the main interlocutor. They also include products that blend AI replies with human-curated content, and require disclosure for the AI-generated parts. A robot voiced live by an operator is therefore treated differently from one answering with a language model. Mixed set-ups need careful thought. Our guide to teleoperated and autonomous robots explains how these modes look to an audience.

The guidelines' examples of deepfakes include a realistic synthetic avatar of a company CEO addressing staff, and AI-generated depictions of a celebrity in promotion. A robot or screen that uses a convincing likeness or cloned voice of a real person may therefore need a deepfake label from the deployer. A fictional robot character with an invented voice is, on the guidelines' examples, much less likely to be a deepfake.

For generated content, the Commission published a final Code of Practice on Transparency of AI-generated Content in June 2026, together with a set of EU icons for labelling. The Commission states that about 190 organisations had signed it by the end of July 2026. The code covers marking and labelling under Article 50(2) and 50(4), not the disclosure duty for interactive systems.

Enforcement and penalties

National market surveillance authorities enforce Article 50, with the AI Office competent for some systems built on general-purpose AI models. According to the Commission's guidelines, breaches can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower of the two applies. Anyone may lodge a complaint with an authority.

What this means for event organisers

• Ask how the robot discloses that it is AI. In your brief to a robot supplier, ask for the opening line, its languages and how it handles the question "are you a person?".
• Plan for every new visitor, not just the first. At a busy trade-show booth, a spoken greeting combined with visible signage on or near the robot reaches people who arrive mid-conversation.
• Do not rely on the robot looking mechanical. If children or the general public may approach, the Commission's reading makes the obviousness exception hard to use.
• Treat cloned voices, real-person likenesses and emotion analysis as separate decisions. Each adds deployer duties, and some emotion recognition is prohibited.
• Be clear about who is speaking. If an operator voices the robot at some moments and AI at others, agree in advance how visitors will be told.

Frequently asked questions

Does Article 50 apply to a humanoid robot that talks to visitors?

In most cases, yes. The European Commission's guidelines of July 2026 list humanoid robots, cobots and robotic companion pets as examples of AI systems that interact directly with people. If the robot uses AI to hold a conversation, its provider must ensure people are told they are dealing with AI, unless that is obvious to the whole foreseeable audience. That is difficult to show in public spaces with children present.

When did the Article 50 transparency rules start to apply?

Article 50 has applied since 2 August 2026. The Digital Omnibus on AI added a transition only for the machine-readable marking duty in Article 50(2): generative systems already on the market before 2 August 2026 have until 2 December 2026. The duty to tell people they are interacting with AI was not postponed. Content generated before 2 August 2026 does not need retroactive labelling.

Is a notice in the rental contract or the robot's manual enough?

According to the Commission's guidelines, no. A disclosure that appears only in terms and conditions, documentation or a URL is insufficient on its own. The information must be clear, noticeable and given at the latest when each person starts interacting. For a speaking robot, the guidelines suggest a spoken statement at the start, possibly combined with visual labels and reminders in longer interactions.

Does the rule apply if a human operator speaks through the robot?

The guidelines exclude mediated interaction, where a human relays or reviews AI output and remains the main interlocutor. A robot voiced live by an operator is therefore treated differently from one replying through a language model. Where AI replies and human input are mixed, the guidelines require disclosure for the AI-generated parts. This is general information, not legal advice, so check specific set-ups with an adviser.

What are the penalties for breaching Article 50?

According to the Commission's guidelines, providers and deployers that breach Article 50 can be fined up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower amount applies. National market surveillance authorities enforce the rules, and any person can lodge a complaint. Fines must take account of the seriousness and duration of the breach and other circumstances.

How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-11-03. Spotted an error? Write to info@humanoid4rent.com.