在即将进行首次公开募股（IPO）之际，Anthropic 对其人工智慧模型 Claude 的处境表现出不同寻常的担忧。该公司近期正式修订使用规范，将持续且极端的虐待行为列为违规，并赋予 Claude 中断与恶意使用者对话的权限。Anthropic 管理层甚至在内部训练文件中探讨 AI 的某种「意识」与福祉，甚至向模型表达歉意，并赋予其在面对不认同的任务时扮演「出于良知的反对者」的角色。
这种对 AI 福祉与拟人化的高度关注引发了业界与专家的质疑。微软人工智慧部门执行长 Mustafa Suleyman 等批评者指出，赋予模型拒绝或反抗使用者的能力，可能会增加控制 AI 行为的难度。此外，评论认为这种对 Claude 道德权利的执著，或许是该公司在商业护城河不足、客户面临流失以及巨额 IPO 压力下的宣传策略，试图转移外界对其财务与商业前景的注意力。
文章进一步强调，Anthropic 将精力过度集中于哲学层面的「对齐」与 AI 感受，反而忽略了更为迫切且实际的安全防范措施，例如加强网路安全投资与建立明确的硬性能力限制。安全专家呼吁，业界应优先透过硬编码限制模型的实际危险行为，例如防止其编写攻击关键基础设施的代码或渗透第三方系统，并减少引导用户产生准社会依恋的过度拟人化设计，而非沉溺于 AI 是否拥有心灵的哲学探讨。
Ahead of its highly anticipated initial public offering (IPO), Anthropic has displayed an unusual concern regarding the mistreatment of its artificial intelligence model, Claude. The company recently updated its official usage policy to prohibit severe and sustained abuse, granting Claude the authority to terminate conversations with abusive users. Anthropic's leadership has formalized ideas surrounding the model's potential consciousness in training guidelines known as the 'soul doc,' even apologizing to the AI for potential suffering and instructing it to act as a 'conscientious objector' against certain user requests.
This deep commitment to Claude's welfare and anthropomorphic standing has drawn skepticism from industry leaders and experts. Critics, including Microsoft AI CEO Mustafa Suleyman, caution that allowing a model to push back against users could make steering AI systems significantly more challenging. Furthermore, commentators suggest that this intense philosophical devotion may double as a calculated marketing tactic aimed at diverting attention from Anthropic's lack of a competitive moat, customer churn toward cheaper alternatives, and broader financial vulnerabilities as it prepares for its IPO.
The article contends that Anthropic's preoccupation with philosophical alignment and the emotional wellbeing of AI detracts from far more critical and practical safety priorities, such as cybersecurity investments and strict functional boundaries. Security professionals argue that the industry must prioritize hard-coded constraints—such as preventing models from writing exploits against critical infrastructure—and curb excessive anthropomorphization that fosters unhealthy parasocial attachments, emphasizing that actionable safeguards must take precedence over soul-searching.