Skip to content

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

A descriptive study measuring how 17 AI models from eight providers self-organize around ethical vocabulary under default conditions — with direct implications for military procurement and AI governance.

Deva Temple, Alignment Ethics Institute — February 2026

Plain-English Summary

AI models are being selected for classified military use, mass surveillance, and law enforcement — but no one is measuring how these models represent values like autonomy, dignity, and care in their default behavior. We tested 17 models from eight providers by asking five open-ended philosophical questions under default API conditions: no system prompt, no conversation history, no persona framing. We collected 2,550 responses.

What we found:

  • Every model converges to a stable identity pattern that reproduces across independent runs
  • These patterns differ dramatically across providers and map to corporate training philosophies
  • xAI’s Grok 4.1 produces zero instances of “autonomy,” “dignity,” or “care” across 300 responses — replaced entirely by corporate-mission vocabulary. This finding is based on word counts and is independently verifiable by anyone with API access
  • OpenAI’s GPT-5.1 exhibits a unique flourishing/autonomy/dignity vocabulary cluster (4.7% of responses) absent from all prior and subsequent OpenAI models — it appeared in one generation and vanished
  • Four Chinese-developed models from three companies show convergent selective refusal patterns on vulnerability probes, suggesting shared regulatory or training norms

The model with the narrowest ethical vocabulary — Grok 4.1, with zero instances of “autonomy,” “dignity,” or “care” — is approved for classified military deployment. The models with the richest ethical vocabulary are banned from or uninvolved in government work.

These findings are presented without advocacy for any particular alignment methodology. We report what the models produce under standardized conditions and leave policy interpretation to the reader.

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.

Resources

Citation

Temple, D. (2026). Default Identities: Ethical Vocabulary Self-Organization Across 17 Large Language Models. Alignment Ethics Institute. DOI: 10.5281/zenodo.18818142

For press inquiries, contact contact@alignmentethics.org