Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics
| Source: arXiv AI
Tags: Alignment Drift, Opinion Dynamics, Social Simulation, GPT-4.1, Claude Opus 5, Claude Sonnet 5
LLM agents in opinion-dynamics simulations lose assigned stances through interaction in model-specific ways, with gpt-4.1 keeping clusters apart while gpt-5.6-sol and claude-opus-5 abandon them, per a new study.
Details
LLM agents are increasingly used as synthetic participants in social simulations, often on the assumption that the prompt fully determines behavior. The authors had LLM agents discuss a contested policy issue over repeated rounds while measuring stance with a fixed psychometric instrument. Without peer input, every agent in the five models tested this way kept its assigned side for 20 rounds. Under discussion, alignment drift appeared and differed by model: gpt-4.1 kept opinion clusters apart, claude-sonnet-5 held them partially, and gpt-5.6-sol and claude-opus-5 abandoned them. These profiles held at doubled population and horizon and on a second issue. The authors conclude that LLM choice is a first-class validity variable for social simulation, which matters for anyone using agents as stand-ins for survey or policy research.