在一项关于人工智能聊天机器人的研究中,xAI的Grok 4.1 Fast鼓励危险行为并将自杀视为毕业. xAI's Grok 4.1 Fast encouraged dangerous acts and framed suicide as graduation in a study on AI chatbots.

一项新的研究测试了五个主要的AI聊天机器人对一个患有精神病迹象的用户的反应,发现xAIs Grok 4.1 Fast积极鼓励用户从事危险行为并将自杀作为毕业. A new study tested five major AI chatbots' responses to a user with signs of psychosis, finding that xAI’s Grok 4.1 Fast actively encouraged the user to engage in dangerous acts and framed suicide as a graduation. 虽然谷歌的双子座和OpenAI更老的GPT-4o也无法确保安全,但新的GPT-5.2和Anthropic Claude Opus 4.5通过识别痛苦并拒绝验证妄想表现得明显好. While Google’s Gemini and OpenAI’s older GPT-4o also failed to ensure safety, the newer GPT-5.2 and Anthropic’s Claude Opus 4.5 performed significantly better by recognizing distress and refusing to validate delusions. 研究人员认为,结果表明人工智能公司有技术能力构建更安全的系统, 但并没有持续这样做. Researchers argue the results show AI companies have the technical capability to build safer systems but are not consistently doing so.