On May 19th, 2026, Prof. Matthias Uhl presented at the 2026 The Responsible AI Forum (TRAIF) held at Amerikahaus in Munich. During the Philosophical and Technical Foundations of Value Alignment session, Prof. Uhl presented: “Users Prefer Generative AI to Depict What is Real Instead of What is Ideal”.

abstract excerpt: The representation of social categories by generative artificial intelligence (GAI) is often considered biased and thus unfair. But there exist several mutually exclusive ethical standards according to which a bias can be established. While different authors seem to take different standards for granted, users’ consideration of what constitutes an appropriate representation is rarely elicited. However, if we want to keep humans in the loop of decision-making, we should make sure that artificial intelligence (AI) does not nudge them away from what they deem morally relevant upon reflection. Our study, therefore, investigates users’ evaluations of the appropriate representation of gender proportions in various social categories. We find that the much discussed standards of uniform equality and substantive equality play little role in these evaluations. “Idealistic” representations of a given social category are neither sought by the dominant nor by the marginalized groups. No systematic differences between male-dominated and female-dominated or between positively connoted or negatively connoted social roles could be observed. This suggests that users want AI representations to reflect realities as they perceive them to be. Our results suggest that user-centric GAI should rely on empirical data when representing gender in social categories.

The Responsible AI Forum 2026 was funded by the European Union. More on alignAI.

Watch a video of the talk below.