What we decided, and what the tests left of it
In short
In the journal, the founders of Aura51 explain the decisions behind building AI avatars for the stage. A new post appears when we make a fundamental decision or a test proves us wrong. The first two are about dissent in the briefing and about why this website states what does not work yet.
Last updated 2 October 2026
- Article · 2 October 2026
Dissent on demand: does an AI avatar make a good devil’s advocate?
We think a panel without an opposing voice is wasted, because only an objection forces the guests to give reasons. That holds for role-played dissent too. When everyone on the panel agrees, an AI avatar briefed with a position can play devil’s advocate. It does not replace a real opponent, since a 2001 study found role-played dissent to be weaker.
- Article · 2 October 2026
Why we publish what doesn’t work yet
Because a flattering number gets caught at the first event, and then nobody trusts the correct ones either. So each capability of our AI avatars shows how far it’s proven, even when it has only been measured in the lab. If an answer drops the caveat, the website won’t build.
01First posts
Two decisions someone could hold against us
We started with two posts that, at first glance, argue against our own interest.
In the first, we advise organizers against our AI avatar under one condition. If you know someone who genuinely disagrees, invite that person. The post explains why we let the AI avatar push back anyway, and the answer lies in the difference between a working group and a panel.
The second explains why every capability on this website states how far it has been proven, even where a round number would look better. It opens with a test that went wrong.
02From the tests
From “too long” to “too short”
Not one answer shorter than 210 characters. That was the AI avatar when we began testing live. A technical limit decided when it stopped, not the question. We set a tighter limit of 240 characters and told it to hand back the floor sooner.
After the next test, we tied the limit to the kind of question, because a fixed limit is always wrong for one kind of question: too loose for a follow-up, too tight for an explanation someone asked for. Say “kurz” (briefly) or ask for a “Schlusswort” (closing word) and the answer is short; ask for “ein Beispiel” (an example) and it is longer. The moderator does not have to set anything. Later, the test group found its answers too brief all the same: too many ended right at the limit for ordinary questions. So since then, that limit grows with the length of the question.
The limit was still too low. In the following run, its answers to being called by name all hit the limit again. After another correction, the AI avatar ended its answers on its own in the run after that. Asked to explain something in more detail, it spoke 516 characters.
03Rules
Rules you can hold us to
The posts are written by the founders, Oliver Schibli and Martin Radelfinger. Four rules apply to each one.
- Every post carries the date it was last changed.
- We announce an opinion before we argue for it. That is why the post on dissent opens with “We think …”.
- Studies and figures from third parties come with their primary source below the post.
- A scene we make up is labeled as an example in the text.
04Still open
The post that is still missing
After every pilot event, a report appears here, with the setup and the figures from the room.