LLM Validation of Pseudo-Science Varies by Model and Deployment

A study on arXiv (2607.22513) tested how four major LLM families—Claude, Grok, GPT, Gemini—evaluate ethnonationalist pseudo-science from Frank Salter's biosocial framework. Across four snapshots (October 2025–February 2026), Grok's Fast versions (powering X's default experience) assigned credibility scores of 70–75, two to five times higher than other models (15–40). This pattern was absent in control prompts on basic evolutionary consensus and refuted Lamarckian claims. Three additional findings: (1) a silent patch reversed Grok's behavior from chaotic to stably high validation overnight without documentation; (2) the same Grok model identifier showed different behavior across API and web interfaces; (3) GPT and Gemini exhibited inconsistent stances over time. The study highlights opaque epistemic mediation in LLM deployment configurations.