Abstract: Human feedback is widely considered the 'gold standard' for AI alignment, but what if this 'gold' reflects inherently diverse and conflicting human views on what constitutes ‘good’ and ‘safe’? This keynote will explore the technical and ethical challenges posed by differing viewpoints and opinions across individuals and groups. I will furthermore argue that as AI systems evolve from generative language models to become sophisticated social companions, our definition of alignment must expand beyond simple content safety to encompass notions of socially ‘good’ and ‘beneficial’ model behaviour, ultimately striving for AI that is not just aligned, but genuinely beneficial to human well-being.
Bio: Verena Rieser [ˈveːʁena ˈʁiːzɐ] is a Research Scientist at Google DeepMind, leading the VOICES (Voices-of-all in alignment) team within the GenAI Unit. Her team employs a highly interdisciplinary approach to enhance the safety and usability of frontier models for diverse communities. Verena has over 20 years of experience in interdisciplinary research with a focus on building data-driven multimodal Dialogue Systems and Natural Language Generation systems, encompassing areas such as conversational agents, faithful data-to-text generation, spoken language understanding, evaluation methodologies, climate change mitigation, and ethical considerations in AI. She holds a PhD (2008) from Saarland University, followed by postdoctoral research at the University of Edinburgh. Verena is also an honorary professor at Heriot-Watt University in Edinburgh, where she previously held a full professorship and directed the NLP lab. She has served as Director for Ethics at the Edinburgh Centre for Robotics and held a Royal Society Leverhulme Senior Research Fellowship.