"I apologize for my actions": Emergent Properties and Technical Challenges of Generative Agents

This work explores the design, implementation, and usage of generative agents towards simulating human behaviour. Through simulating (mis)information spread, we investigate the emergent social behaviours they produce.Generative agents demonstrate robustness to (mis)information spread, showing realistic conversational patterns. However, this robustness limits agents’ abilities to realistically simulate humanlike information dissemination. Generative agents also exhibit novel and realistic emergent social behaviours, such as deception, confrontation, and internalized regret. Using deception, agents avoid certain conversations. Through confrontation, an agent can verify information or even apologize for their actions. Lastly, internalized regret displays direct evidence that agents can internalize their experiences and act on them in a human-like way, such as through expressing remorse for their actions.We also identify significant technical dynamics and other phenomena. Generative agents are vulnerable to produce unrealistic hallucinations, but can also produce confabulations which fill in logical gaps and discontinuities to improve realism. We also identify the novel dynamics of "contextual eavesdropping" and "behavioural poisoning". Via contextual eavesdropping and behavioural poisoning, agent behaviour is altered through information leakage and sensitivity to certain statements, respectively.