Does an AI avatar at an event have to be labeled?

01The law

An invented character has to identify itself too

Suppose a CEO addresses the company town hall in the morning as an AI avatar, with his face and voice. In the afternoon, Albert, an invented character, sits on a conference panel. Under the European Commission’s guidelines, the CEO avatar can be a deepfake. In our reading, Albert is not.

The guests debating Albert on stage still have to be told that an AI is answering. Article 50 of the EU AI Act, as amended in July 2026, requires this for every AI avatar that talks to people.

• Aug 2, 2026Article 50 appliesArticle 113 of the AI Act
• €15 millionmaximum fineor 3% of worldwide annual turnover; for SMEs, whichever is lower
• Dec 2, 2026marking for older systemsfor systems already on the market before Aug 2, 2026; Article 111(4)

• Paragraph 1 · provider

  Disclose that an AI is speaking

  An AI system that interacts directly with people must be designed so that they are informed of it, unless this is obvious to a reasonably well-informed, observant person. The Commission’s guidelines name avatars as an explicit example and read the exception narrowly.

• Paragraph 2 · provider

  Mark image and sound machine-readably

  Providers of systems that generate synthetic images, video, audio or text mark these outputs as artificial in a machine-readable format. That marking is read by machines, not by the audience.

• Paragraph 4 · deployer

  Disclose deepfakes

  Whoever uses a deepfake discloses that the content is artificially generated or manipulated.

Paragraph 1 protects everyone who talks to the AI, the moderator included. Like the guests, the moderator must know by the first exchange at the latest, clearly and in an accessible way (paragraph 5). For those in the room who only listen, a notice is good practice.

02Deepfake

When is an AI avatar a deepfake?

The Commission’s guidelines name three criteria, and only all of them together make a deepfake.

1. 01

   Resemblance

   The content closely resembles its original. The CEO avatar has the CEO’s face and voice.

2. 02

   A real original

   The original actually exists. The CEO does. Albert is modeled on no one, so in our reading the check ends here for him.

3. 03

   False appearance

   The content falsely appears authentic. That applies to the CEO avatar as soon as people could take it for the CEO himself.

For evidently artistic, satirical or fictional works, such as a parody at a gala, a disclosure that does not hamper the work is enough.

03Responsibility

At a conference, who is the provider and who is the deployer?

Take an industry association hosting a conference. In the closing panel, an AI avatar debates three guests. The company that developed it and offers it under its own name is the provider. It makes sure the three guests and the moderator are told (paragraph 1) and marks image and sound machine-readably (paragraph 2).

The deployer is whoever uses the avatar that afternoon. In our reading, what counts is responsibility for the use, not whose hand is on the laptop, and as long as an invented character sits on the panel, Article 50 remains the provider’s business. If the association has its president appear as an avatar instead, it must disclose the deepfake as the deployer (paragraph 4).

04Switzerland

Does the AI Act apply to events in Switzerland?

If an AI avatar from Horgen appears at a congress in Munich or Vienna, Article 50 applies. The AI Act also covers providers and deployers outside the EU once their system’s output is used in the EU (Article 2(1)). At the same congress in Zurich, Article 50 is generally not binding. It can be different if the appearance is streamed into the EU or an organizer based in the EU deploys the avatar (Article 2(1)).

• Feb 12, 2025Federal Council decisionto ratify the Council of Europe’s AI convention and adapt Swiss law where needed
• End of 2026draft for consultationprepared by the Federal Office of Justice

Switzerland has no AI law of its own yet. The Federal Data Protection and Information Commissioner (FDPIC) does state, however, that the existing Data Protection Act applies directly to AI. From this it concludes that anyone communicating with a language model has a legal right to know whether they are talking to a machine. Some data protection experts read the law more narrowly and see no general right of this kind.

That is one more reason we label it in Zurich as well as in Munich. Anyone arguing with an AI avatar on stage should know who they are dealing with, wherever the conference is held.

05Aura51

Albert, Cassandra, Ophelia and Gustav have no real-life original

What applies to Albert above applies to all four. They are entirely virtual, and none is modeled on a real person, living or dead.

Paragraph 1 still applies, because guests on stage talk to them. For the room, the video carries the label “AI avatar” for the entire appearance, at every event, including outside the EU. Before the first call, the moderator announces that an AI avatar is joining the discussion.

06Checklist

What to settle before the event

• Agree with the provider who counts as the deployer and how the video on the screen is labeled.
• Tell the panel guests and the moderator in the invitation that an AI avatar will join the discussion.
• Inform the room at the start, with an on-screen notice or a line from the moderator. It might sound like this: “Joining us today is Albert. He is an AI avatar, and his answers are written by an AI.”
• If the avatar is meant to resemble a real person, disclose the recreation so that the room learns of it too. That is how we read paragraph 4.
• If the event takes place in the EU, have the provider confirm in writing how it meets paragraphs 1 and 2.