Where are the limits of an AI avatar on stage?
In short
Some things the AI avatar leaves out on purpose. It does not speak unprompted or make up numbers. Other things the technology does not allow. The round with it runs in standard German and only with internet. Its first word comes 1.5–2 s after a question, measured live; later if someone tacks on a second question.
Last updated 2 October 2026
01Two kinds
Over five seconds of silence, though it waited as it should
Anyone moderating a round with an AI avatar first notices its limits as silence. After a question the room waits longer than it would for a guest on the panel, and sometimes the moderator stretches that pause without noticing.
During one of our live tests the moderator called the avatar by name and asked for its view, then tacked on a second question right away. It waited for that one to end too, which is how it is built. But speech recognition only closed the second sentence about four seconds later. More than five seconds passed before it said a word.
It had done nothing wrong. The silence came from the technology. Other limits we drew ourselves.
What it will not do, by our choice
- Speak unprompted. It talks when the moderator calls on it or gives it the floor with a button, and briefly after that to follow-up questions.
- Make up numbers or studies.
- Listen in on the audience on its own. By default it hears only the stage microphones. If a room microphone is on the aux send, voices from the floor reach it too. Otherwise the moderator passes a question from the floor on to it.
What it cannot do, because the technology does not allow it
- Answer in less than about 1.5 s.
- Understand or speak Swiss German dialect. It works in standard German; English is not possible yet either.
- Tell reliably, without its name, whether a question is meant for it. After an answer it takes follow-ups; the conversation ends 12 s after the last sentence. We set that time ourselves. After that it needs its name again.
- Work without internet.
02Numbers
Why it would rather name a gap than a number
It does not make up numbers or studies. We decided that, just as we decided it stays silent until called. A language model can produce a plausible figure that exists nowhere. On stage that weighs more than on a screen, because people in the room write down a number once it is spoken, and hardly anyone there can check it.
So the avatar is instructed to name a gap openly, and a wrong number that slips through anyway reaches the moderator’s tablet first, where the moderator can stop it. When it says on stage that it does not know something, it is working as intended. Pass the question to a guest who knows the figure.
03At the event
Which limit affects your event?
Agree with your guests beforehand that the part with the AI avatar runs in standard German. A remark in Swiss German passes it by.
Ask one question per call, then let the pause stand. Every sentence the moderator adds makes it longer, because the avatar waits for that sentence to end too.
Right after its answer, a follow-up without its name often works. But it cannot reliably tell that a “you” question without its name is meant for it; where it matters, for instance with a new topic, say its name.