Long-term social alignment and behavioral adaptation in human–AI interaction

Determine whether AI agents built on large language models, when exposed to sustained interactions with humans over time in dynamic, multi-user environments, develop shared norms, adapt to user values, or exhibit behavioral drift.

Background

The paper surveys emergent behaviors in human–AI interactions, highlighting roles such as companion, catalyst, and clarifier, and noting structural asymmetries between humans and AI agents (e.g., memory persistence and access to broader context). It emphasizes that most existing evaluations focus on short-term outcomes and use human-centric metrics, leaving the mechanisms of agent behavior in hybrid settings underexplored.

Within this context, the authors raise a specific uncertainty regarding the temporal dynamics of agent behavior under prolonged exposure to human users. They question whether agents would converge toward shared norms, align with user values, or drift behaviorally over time—issues that directly affect the long-term social alignment of AI systems operating in multi-user environments.