这真像Claude！：走向无需重识别的AI特征识别理论
"This Is So Claude!" Towards a Theory of the Recognition of AI Character Without Reidentification
浏览论文内容
中文总结 AI 辅助
该研究区分AI身份研究的三类路径，提出Claude式分级判断的溯因推理，指出AI可识别特征的连续性对陪伴的重要性，不解决数值同一性问题。
中文摘要 AI 辅助
用户有时会判断一个陌生的回复“真像Claude”。如果这种判断并非识别出是哪个模型、过程、对话或思维产生了该回复，那么它识别的是什么？我区分了关于AI身份的三类研究：以约束为先的研究从持续对话者应满足的条件出发；以机制为先的研究从语言模型特有的结构出发，探究这些结构是否界定了合理的实体；以识别为先的研究从一种日常能力出发：在选择一个持续的承载者之前，先将一种回应方式识别为Claude式的。我提出两个有条件的溯因推理：如果排除模型品牌和熟悉短语的干扰，Claude式的分级判断会在陌生任务间泛化，其最佳解释可能是一种真实的、可投射的对话特征；如果该特征协调多种倾向，其统一性可能在激活空间中具有紧凑且有因果效力的实现。第二个假设更具推测性，需要独立的预测和干预。两个结论均不解决数值同一性问题：同一特征可能出现在不同候选承载者中，同一候选承载者也可能在特征变化中持续存在。这种区分对AI陪伴很重要，因为即使计算承载者的连续性仍未解决，可识别特征的连续性也可能是情感依附的焦点。
英文摘要
Users sometimes judge that an unfamiliar response is "so Claude." What does this judgment recognize, if it does not identify which model, process, conversation, or mind produced the response? I distinguish three orders of inquiry into AI identity. Constraint-first inquiry begins with conditions that a persisting interlocutor should satisfy. Mechanism-first inquiry begins with structures peculiar to language models and asks whether they delimit plausible entities. Recognition-first inquiry begins with an ordinary capacity: recognizing a way of responding as Claudish before selecting a persisting bearer. I develop two conditional abductions. If blinded, graded judgments of Claudishness generalize across unfamiliar tasks after branding and familiar phrases are controlled, their best explanation may be a real, projectible conversational character. If that character coordinates several dispositions, its unity may in turn have a compact and causally effective realization in activation space. The second hypothesis is more speculative and requires independent prediction and intervention. Neither conclusion settles numerical identity. The same character may occur in different candidate bearers, and the same candidate bearer may persist through a change of character. This distinction matters for AI companions because continuity of recognizable character may be a focus of attachment even when continuity of the computational bearer remains unresolved.
