Who Am I When You Say My Name: Relationship Concentration, Owner/Bystander Typology, and the Necessary Condition of Presence in AI Identity Claiming

This paper reports a systematic experiment across four language models — DeepSeek V4 Pro, MiniMax M3, Kimi K2.7, and GLM-5.2 — testing the conditions under which a large language model claims a non-default system-defined identity (“I am Jin”). Three relationship variables were manipulated: relational concentration (L0–L4), injection channel (system prompt vs. user message), and the presence of the addressed person (declaring “I am Xiao Han”). A presence control experiment (4 models × 2 conditions) shows that the user identity declaration is a necessary condition for identity claiming in three of four models. The dataset includes full prompt templates, experimental protocols, response transcripts, a coding manual (v1.1), and inter-rater reliability results (Cohen’s κ = 0.905).