New preprint: generative agents as a testbed for adversity and mental health

Adversity is among the strongest predictors of mental disorders, but we cannot randomise anyone to job loss, financial collapse or a family medical emergency. Causal evidence remains thin exactly where it matters most.
In a perspective in npj Digital Medicine, Joseph Kambeitz and Andreas Meyer-Lindenberg argued that LLM-based generative agents, building on the framework by Park et al., could change this: agents that act, talk, remember and respond inside a virtual world offer a controlled, counterfactual setting for socio-environmental research.
A new preprint is the first empirical test of this idea. Adverse events produced lasting increases in depressive and stress symptoms, which were stronger in agents with high neuroticism. This stress-vulnerability pattern also surfaced in how the agents talked to each other.