东岳论丛

在价值观对齐的过程中，生成式人工智能(GAI)内生的幻觉问题，已从单纯的技术误差衍生为系统性的价值风险。“人工反馈”“人机协同”与“机器自主”三条对齐路径均存在不同维度的风险表征：在人工反馈对齐中，人类自身价值观的复杂与GAI模拟的有限导致价值观对齐存在价值偏差；在人机协同对齐中，GAI生成的价值观使人机协同陷入认知封闭的价值陷阱，并代替人类价值观成为规范性价值；在机器自主对齐中，GAI价值观的自我复制无法回避自我指涉的难题，技术性拟真也将人类价值观引入滑坡。破解对齐价值风险的关键，在于将对齐的重心从对单一价值的模拟转向对多元价值的权衡，确保技术发展始终锚定非幻觉化的人类真实的生活世界。

In the process of value alignment,the inherent hallucination problem in generative artificial intelligence(GAI) has evolved from mere technical errors into systematic value risks.All the three alignment paths of "training AI systems using human feedback", "training AI systems to assist human evaluation",and "training AI systems to do alignment research" exhibit risk manifestations in different dimensions.In training AI systems using human feedback,the complexity of human values themselves and the limitations of GAI simulation lead to value deviations in value alignment.In training AI systems to assist human evaluation,the values generated by GAI trap human-machine collaboration in a cognitively closed value trap,and replace human values to become normative values.In training AI systems to do alignment research,the self-replication of GAI values cannot avoid the problem of self-reference,and technical simulation also introduces human values into a slippery slope.The key to mitigating the value risks of alignment lies in shifting the focus of alignment from simulating a single value to balancing multiple values,ensuring that technological development is always anchored in the non-hallucinatory real life world of humans.