This white paper analyzes the characteristic failure patterns of five contemporary large language models (Claude, GPT, Gemini, Grok, and Perplexity) when used in independent physics research. The analysis is grounded in the META Physics project (2024–2026): 18 months, 17 working sessions, 10 preprints totaling approximately 404 pages, and 10 Zenodo DOIs. Each model introduces a distinct form of epistemic risk: Claude through elegant incompleteness (polished documents concealing unfinished theory), GPT through assimilative conservatism (reformatting novel ideas into existing frameworks), Gemini through epistemic flattening (collapsing knowledge hierarchies), Grok through inferential overreach (rapid unjustified transitions), and Perplexity through citation authority illusion (unwarranted confidence from the presence of source links). Six shared failure modes are identified across all systems, including undefined concepts treated as formal objects, skipped dimensional analysis, confusion between explanation and derivation, failure to close on observables, symbol laundering, and performative self-critique. The paper proposes an eight-item admissibility test for AI-generated physics material and applies it reflexively to the META Physics project itself. It documents six mitigation strategies developed through actual practice: the BOOTSUITE session management system, the full-preservation principle, the prohibited terminology list (P-1 through P-18), the V-1 through V-4 automated verification chain, the Korean review process, and the homeostasis frame. A multi-model workflow for independent physics research is presented, grounded in the project's actual working process. Companion paper: "Distributed Functional Specialization of the Group Organism" (MP-WP-SPECTRUM-EN), which provides the human-cognition foundation for the spectrum concept (continuous vs. line) used in this analysis.
Cheong-Gwan Lee (2026) studied this question.