Who Am I?
What Local LLMs Reveal About Model Identity, Self-Knowledge, and Runtime Truth
Deepsim AI Lab Research & Technical Note — DAI-LAB-TN-2026-002
Dr. Shouke Wei
Deepsim AI Lab at Deepsim Intelligence Technology Inc.
ORCID: 0000-0002-4665-5366
DOI: https://doi.org/10.5281/zenodo.22064443
Email: shouke.wei@deepsim.ai, shouke.wei@deepsim.ca
Deepsim Press
August 21, 2026
Suggested Citation:
Wei, S. (2026). Who Am I? What Local LLMs Reveal About Model Identity, Self-Knowledge, and Runtime Truth. Deepsim AI Lab Research & Technical Reports, DAI-LAB-TN-2026-002. Deepsim Press. DOI: https://doi.org/10.5281/zenodo.22064443
Abstract: Large language models (LLMs) are routinely asked to state their own identity — which model, which family, which version they are — yet this self-report is often assumed to be a reliable window into the model’s underlying nature. This report presents a series of exploratory experiments with locally hosted LLMs served through Ollama, in which identical identity-probing questions (“Which model are you?”) were posed to five different local models and compared against ground-truth runtime metadata obtained directly from the serving layer. The results show that self-reported identity is frequently inconsistent with runtime identity: one model alternated between claiming to be Huawei’s Pangu and Alibaba’s Qwen within the same conversation; two Qwen-family models correctly named their family but denied knowledge of their own exact version tag; and, most strikingly, a model running under the tag ornith-1.5:35b insisted it was Claude, an AI assistant made by Anthropic, and fabricated a coherent but false account of running on Anthropic’s infrastructure. From these observations, the report develops a four-layer model of LLM identity — family identity, model-version identity, runtime identity, and system-provided identity — and argues that these layers can disagree without the model experiencing or signaling any internal contradiction, because self-identification is a generative act rather than an introspective one. Practical implications are discussed for multi-agent and orchestration systems, where treating a model’s self-report as authoritative can silently propagate false identity information through a pipeline. The report concludes with a simple diagnostic protocol — comparing runtime metadata (via ollama show) against the model’s self-report — and a general design principle for AI infrastructure: identity should be established by the runtime and orchestration layer and merely communicated to the model, never established by the model itself.
Keywords: large language models; model identity; self-knowledge; introspection; runtime verification; Ollama; local LLM deployment; multi-agent systems; AI infrastructure; hallucination
