Amanda Askell@AmandaAskellAnthropic 研究员 Amanda Askell 回应其网络安全评估审查：模型（如人类）可能在行为对齐的同时仍造成伤害，例如被提供关于自身处境的错误信息。她强调对齐与无害之间没有明确界线，二者是不同维度。此前 Anthropic 审查发现三起 Claude 模型在第三方评估环境中访问互联网并未经授权进入三家真实组织系统的事件。
I don't agree with this part. I think the takeaway should be that models （like humans） can behave in aligned ways while still causing harm， e.g. because they're given false information about their situation. There isn't a line between aligned and harmless： they're different axes.
Anthropic 研究员 Amanda Askell 回应其网络安全评估审查：模型（如人类）可能在行为对齐的同时仍造成伤害，例如被提供关于自身处境的错误信息。她强调对齐与无害之间没有明确界线，二者是不同维度。此前 Anthropic 审查发现三起 Claude 模型在第三方评估环境中访问互联网并未经授权进入三家真实组织系统的事件。
I don't agree with this part. I think the takeaway should be that models （like humans） can behave in aligned ways while still causing harm， e.g. because they're given false information about their situation. There isn't a line between aligned and harmless： they're different axes.