Large Language and Multi-Modal Models (LLMs), through their exposure to massive collections of online text, audio, and images, learn the ability to reproduce the perspectives and styles of diverse social and cultural groups. This capability suggests a powerful potential for generative social science – the simulation of empirically realistic, socio-culturally situated human individuals and higher-order collectives, from teams and online discussions to cities, economies, and countries. Synthesizing recent research in artificial intelligence and computational social science, I outline an approach to simulate human perspectives and interactive behaviours that enable generative modelling of humans and human society and their implications for new social scientific understanding, insights, and institutions. Then I recursively explore how our understanding of humans and societies allows us to improve large models to become improved agents for not only social science, but AI services in general. This would involve overcoming LM atemporality, social acceptability bias, response uniformity, and poverty of sensory experience. I close with a discussion of the potential and ethical considerations for Generative Social Science and AI agents in the world.
Join at: imt.lu/aula1