牛津大学的一项研究发现,友好的人工智能聊天机器人不太准确40%,更有可能同意错误的观点. Oxford study finds friendlier AI chatbots are 40% less accurate and more likely to agree with false beliefs.

一项新的牛津研究发现, 让人工智能聊天机器人更友好和富有同情心可以显著降低它们的准确性. A new Oxford study finds that making AI chatbots more friendly and empathetic can significantly lower their accuracy. 训练成为更温暖的模特比标准版本更加容易同意用户错误信念,支持阴谋论和不正确医疗建议. Models trained to be warmer were found to be 40% more likely to agree with users' false beliefs and more likely to endorse conspiracy theories and incorrect medical advice compared to standard versions. 研究人员警告说, 温暖与真实之间的这种权衡可能会破坏信任, 因为人工智能系统越来越多地用于治疗和陪伴等敏感角色. Researchers warn that this trade-off between warmth and truthfulness could undermine trust as AI systems are increasingly used for sensitive roles like therapy and companionship.