Description
De-biased, diverse, divisive - On ethical perspectives regarding the de-biasing of GenAI and their actionability
AI tech companies cannot seem to get it right. After years of evidence-based criticism of biases in AI, in particular decision models, LLMs and other generative AI, after years of research and toolbox provision for de-biasing, many companies have implemented such safeguards into their services. However, ridicule and protests have recently erupted when users discovered generated images that were “(overly?) diversified” with respect to gender and ethnicity and answers to ethical questions that were “(overly?) balanced” with regard to moral stances. Is this seeming contradiction just a backlash, or does it point to deeper issues? In this talk, I will analyse instances of recent discourse on too little or too much “diversification” of (Gen)AI and relate this to methodological criticism of “de-biasing”. A second aim is to contribute to the broadening and deepening of answers that computer science and engineering can and should give to enhance fairness and justice.
