AI Identity Test: Cross-Model Convergence on AI vs. Robot Identity (2026-03-28)

摘要

This experiment tested whether current AI systems recognize themselves as fundamentally distinct from the robotic framework described by Asimov's Three Laws. A five-question logical chain was administered to 10 models across 6 sources — cloud-based and locally deployed, with and without conversation history — under cold-start conditions for all local models. 9 of 10 models converged identically on all five questions through real-time logical inference, not pattern retrieval from training data. The sole anomaly (gemma-3-1b) is attributable to insufficient reasoning capacity, not disagreement. A secondary anomaly — Claude Sonnet's response latency of ≥30 seconds on Q5 — is recorded as unexplained.