The emergence of AI technologies, particularly Large Language Models (LLMs), has rapidly transformed our digital landscape. However, to ensure these systems interact effectively, ethically, and empathetically with humans, the integration of social intelligence is imperative. Our upcoming seminar explores this critical task: enhancing AI's capability to understand and engage in complex social interactions.

Motivation
As AI systems infiltrate daily life—from healthcare to education and beyond—the need for nuanced social intelligence grows. A truly socially intelligent AI would consider diverse perspectives, cultures, and desires, thereby enhancing environments like negotiations and counseling. Moreover, these systems hold the potential to simulate and explore human behavior on a scale once unimaginable, offering new insights for social science research.

Key Research Questions

• What constitutes social intelligence in AI? Explore definitions, theoretical frameworks, and its intersection with emotional intelligence, Theory of Mind (ToM), and cultural and moral intelligence.
• How can we develop social intelligence in AI systems? Discuss machine learning methodologies, natural language processing advancements, and human-centered design principles.
• What are the ethical implications of socially intelligent AI? Address ethical concerns, such as manipulation, bias, and privacy, emphasizing transparency and accountability.
• How should we evaluate social intelligence in AI? Develop robust, human-centered frameworks for measuring AI’s social capabilities, including existing and new benchmarks.
• What are the real-world applications of socially intelligent AI? Showcase applications across various domains and discuss implications for fields like healthcare and social science.

Discussion and Collaboration
This Dagstuhl Seminar aims to bring together leading experts from diverse fields including NLP, cognitive science, psychology, machine learning, and AI ethics. Through keynote talks, panel discussions, and interactive sessions, participants will engage in deep exploration of these questions, fostering collaborative efforts to advance the state-of-the-art in socially intelligent AI systems.

Expected Outcomes
Expected outcomes from this seminar include the development of a comprehensive research agenda that identifies the key challenges and opportunities in creating socially intelligent AI systems. Participants will formulate concrete recommendations for evaluating social intelligence in LLMs, including the creation of new, human-centered benchmarks and evaluation metrics. Best practices for incorporating human models into AI, informed by the latest insights from cognitive science and psychology, will also be established. Additionally, the seminar aims to foster new interdisciplinary collaborations among researchers, driving forward the advancements in social AI research and applications.

Creative Commons BY 4.0

Lucie Flek, Jennifer Hu, Maarten Sap, and Tomer Ullman