ºÚÁÏ´«ËÍÃÅ

Artificial intelligence (AI) is increasingly being used to support human decision-making – for example, in medicine, business and public administration. However, even with high-quality data and powerful models, modern AI systems cannot guarantee complete certainty. Uncertainties arise both from flaws in the underlying data (e.g. incompleteness or inaccuracy) and from model uncertainty. If these uncertainties are not explicitly recorded and communicated, there is a risk of unreflective (lack of) trust in AI-based recommendations and, consequently, of erroneous decisions.

The project, funded under Start-up Grant A by the University of Ulm, aims to systematically highlight uncertainties in AI-supported decision-making scenarios and present them in a user-centred manner. The focus is on the development and evaluation of so-called uncertainty-aware AI approaches that support decision-makers in better contextualising AI recommendations and using them in a more reflective manner.

The aim of the project is to develop methods for model-agnostic quantification of data and model uncertainty and to devise suitable mechanisms for communicating these uncertainties in a comprehensible and effective manner. Particular emphasis is placed on the design of human-AI interaction: the research examines how different forms of uncertainty disclosure influence decision-making behaviour and to what extent they can help reduce incorrect decisions.

The approaches developed will be implemented as prototypes and evaluated in empirical user studies. In particular, the research will analyse how the explicit disclosure of uncertainties affects trust, decision quality and error rates in AI-supported decision-making scenarios.

Key research questions are:

1. How can uncertainties in AI systems, resulting from data and model uncertainty, be quantified in a model-agnostic manner?
2. How can these uncertainties be communicated in a user-centred manner to promote reflective AI-supported decisions and reduce decision errors?

Funding body: University of Ulm (ProTrainU, Start-up Funding A)

Project duration: until 2025