What does an AI avatar on stage record, and where is it kept?

01The transcript

What the room does not hear can still be in the transcript

An example: a guest on the panel says something to the woman next to him that is not meant for the room. His microphone is pulled down, and nobody in the room hears it. The remark can still be in the transcript.

That comes down to the path the sound takes. The moderator’s microphone reaches us on a channel of its own; the other stage microphones are routed by your sound engineer from the mixing desk to our laptop over an aux send. That send is taken pre-fader. A microphone pulled down for the room keeps feeding it.

By default, it carries only stage microphones. The AI avatar then picks up no voices from the audience. If a room microphone is on the send, questions from the floor reach the transcript too. If the moderator instead repeats a question aloud or types it on the tablet, the voice of the person asking stays out of the transcript.

02The data path

Where a sentence from the stage goes

Two things reach us before the event starts. At the event, three services in the cloud then come into play.

1. 01

   Briefing

   Topic, position, and the participants with names and job titles. So there is personal data before anyone speaks.

2. 02

   Test recording

   A short recording before the event so the AI avatar recognizes its name. Whoever says the name on it leaves a recording of their own voice.

3. 03

   Sound

   At the event, it goes live from our laptop to a speech recognition service in the cloud.

4. 04

   Transcript

   The service sends back the text, separated by speaker. The transcript is stored on our laptop. To compose an answer, the text goes to a language model in the cloud, together with the briefing.

5. 05

   Answer

   The answer goes as text to an avatar service in the cloud, which turns it into picture and voice for the screen.

03The law

A name next to every sentence: what the law makes of it

Because the transcript puts a name next to every sentence, what a guest says is personal data, just like the name and job title in the briefing. Under both the FADP and the GDPR, that covers any information relating to an identified or identifiable natural person (Art. 5(a) FADP; Art. 4(1) GDPR).

The controller is whoever decides on the purpose and means of the processing. Whoever processes data on the controller’s behalf is a processor (Art. 5(j) and (k) FADP). You invite the guests and decide that an AI avatar takes part. We run the technology and pass the data on to the services. Processing on someone’s behalf requires a contract or a statutory basis (Art. 9(1) FADP; Art. 28(3) GDPR).

The three cloud services are recipients of the data. If data goes abroad, separate rules apply (Art. 16 FADP; Art. 44 GDPR for transfers outside the EU and EEA).

04Information

What your guests should already read in the invitation

The controller must inform the people concerned, as a rule when collecting their data (Art. 19 FADP; Art. 13 and 14 GDPR). So the roles need to be settled before the invitation goes out.

The invitation to your guests on stage should cover four points.

• That an AI avatar takes part in the discussion, and which microphones it listens to.
• Who transcribes and where the text goes.
• Who is responsible and how to reach them, for what purpose and on what legal basis the data is processed, how long it is kept, and what rights the guests have.
• Which service providers receive data and, if data goes abroad, to which country and with which safeguards.

The EU AI Act does not require the invitation itself. For an event in the EU, however, the provider must make sure that guests and moderator know an AI is taking part by the time they first speak with it, unless that is plain to see (Art. 50(1) and (5)). In Switzerland, Article 50 is generally not binding, and the notice remains good practice there. The first of the four points in the invitation is the simplest way to give it.

05Before the event

What is settled and what we clarify with you

Before your event, we clarify with you when the transcript, briefing and test recording are deleted.

This is how the paragraph in your invitation could read. We fill in or delete the square brackets together before the event.

“An AI avatar takes part in the discussion and listens to the microphones on stage [and the room microphone]. A cloud service transcribes what you say, and the transcript shows your name next to it. The transcript is kept on Aura51’s laptop; to compose the AI avatar’s answers, the text goes to a language model in the cloud. The controller is [name, contact]. Data is received by [services, country]. It is deleted after [period].”