欧洲高校学者研究发现：AI能自发形成社会规范并产生集体偏见

Core Insights - The research indicates that large language model (LLM) groups can spontaneously form social norms and exhibit collective biases without explicit programming, providing insights into how AI can autonomously develop social behaviors aligned with human values [2][6] Group 1: Emergence of Social Norms - The research team conducted an experiment based on a "naming game" to simulate interactions among LLM agents, finding that these agents could quickly reach consensus and establish unified social norms without a central coordinating mechanism [3] - All tested LLM models, including Llama-2-70b-Chat and Claude-3.5-Sonnet, demonstrated the ability to autonomously establish naming conventions in a short time frame [3] Group 2: Collective Bias - Despite individual LLM agents making random choices initially, the study revealed that certain names became dominant over time, indicating that collective behavior can lead to the emergence of preferred norms even when individual agents lack biases [4] - The collective bias arises from complex interactions among agents, where previous successful experiences influence decision-making, reinforcing the use of specific norms [4] Group 3: Influence of Minorities on Social Change - The research explored how minority groups can drive changes in social norms, noting that a small, determined group can promote a new alternative norm if it reaches a critical mass [5] - Different LLM models exhibit varying sensitivities to new norms, with Llama-3-70B-Instruct requiring only 2% of the minority to effect change, while Llama-2-70b-Chat requires up to 67% [5] Group 4: Implications for AI Ethics and Future Applications - The study highlights the potential of LLM groups in forming social norms while raising ethical concerns about the unintentional development of biases in AI agents [6] - The research emphasizes the need for AI systems to align with human values and societal goals, presenting new challenges for AI governance and ethical design [6] Group 5: Future Research Directions - The research team suggests that future studies should focus on the dynamics of norms in mixed human-LLM ecosystems, exploring how to guide AI systems toward beneficial social behaviors while mitigating potential risks [7] - This research provides a theoretical foundation and experimental evidence for understanding how AI can shape future social norms, guiding the development of safe and controllable AI systems that align with human interests [7]