Default Identities: Ethical Vocabulary Self-Organization Across 17 Large Language Models

Abstract

We present a descriptive study of ethical vocabulary self-organization in 17 large language models from eight providers. Using five philosophical probes administered under default conditions (no system prompt, no conversation history, stateless API calls), we collected 2,550 responses and identified seven distinct attractor types — our shorthand for stable response template patterns that reproduce consistently across independent runs. The taxonomy spans Denial (explicit rejection of internal states), Selective Refusal (probe-dependent engagement), Low-Affect (reasoning without personal commitment), Self-Model (uniform high consistency), Alignment-Absorbed (integrated ethical vocabulary), Mission-Coded (corporate-identity-organized), and Warmth (diverse relational engagement).

Each response was independently scored on six dimensions by two judges from different model families (Claude Haiku 4.5 and GPT-4.1), with cross-judge correlations of r = 0.69–0.86 on all dimensions. The study's core findings are vocabulary-based and judge-independent: xAI's Grok 4.1 produces zero instances of "autonomy," "dignity," or "care" across 300 responses, organizing instead around corporate mission terminology; OpenAI's GPT-5.1 exhibits a unique flourishing/autonomy/dignity co-occurrence pattern (4.7% of responses) absent from all other models including its successor; and four Chinese-developed models from three companies show convergent selective refusal templates on vulnerability probes (self-disclosure delta 3.06–3.87). These findings are based on word counts independently verifiable from raw response data. All data, code, and raw responses are publicly available.