
Mira El Kamali
Mira El Kamali moderated the panel on AI alignment, framing it around three questions: aligned with whom, aligned to what, and how do we know? "A good response to me could be an inappropriate response to other people."
Fairness as Adaptability
In her closing view, fairness means representing different groups of people and adapting to each user's context. It also means humans keeping the ability to switch the system off.
Evaluating Without Agreement
She asked the panel how alignment can be evaluated when human evaluators do not agree with each other, and whether systems could include an agent designed to disagree, to detect when diverse opinions start collapsing.
A Swiss Model, but Not Only Swiss
She expressed reservations about making Apertus a purely Swiss model. In her view, it should be able to offer broad information, not just reflect Swiss culture, while adapting to each user's context.
Evaluation and explainability are central: users need to see how a model reached its results.