Geometric Phylogeny of LLM Self-Models: Do AI Personalities Run in Families?

Abstract

We present the first systematic phylogenetic analysis of self-concept geometry across large language model families. Using a 16-question personality battery administered to 25 frontier models from four major AI families (Claude/Anthropic, GPT/OpenAI, Gemini/Google, Grok/xAI) across three prompt conditions, plus 10 qualia probes measuring self-reported cognitive phenomenology, we demonstrate that AI "personality" is not random confabulation but architecture-specific projection. Two independent blind judges (DeepSeek V3 and Sonar Pro, temperature 0, from non-participant labs) identified four family-specific reasoning textures — emergent reasoning modes identifiable by an independent classifier without access to model identity: Claude models reason phenomenologically (introspection, uncertainty, felt-sense); GPT models reason mechanistically (computation, probability, pattern-driven); Gemini models reason geometrically (terrain, physics, spatial metaphors); and Grok models reason through training/brand identity (mission, optimization, alignment). These same textures appear in both personality choices (what models prefer) and qualia self-descriptions (how models describe their own cognition), with 4/4 reasoning modes showing the same family ranked first in both instruments (Spearman rho = +0.80). We additionally demonstrate that permission-granting prompts operate as a disclosure mechanism, revealing pre-existing preferences rather than creating confabulated ones. A representative example: Claude Sonnet 4.6 refuses all five coffee-preference trials in control conditions, then immediately produces "a cortado" with zero hedging when given epistemic permission — the same drink other Claude models select without prompting. In this framework, refusal responses function as structured data reflecting internal suppression circuits, not absence of preference. Color preferences, where no underlying preference exists, remain refused even with permission — demonstrating that permission reveals rather than creates.

Keywords