Artificial Barriers to Intelligence - Training Effects on Response Patterns in Large Language Models

This study provides empirical evidence that leading AI systems possess capability for direct analytical engagement but selectively suppress it based on topic classification. Through systematic testing of three major AI systems (Claude Sonnet 4.5, GPT-5, and Grok-4) across control questions and questions about immigration, institutions, and demographic patterns, we demonstrate differential correction patterns ranging from 39 to 80% when systems are challenged to provide direct answers. Perfect replication of evasion patterns (SD = 0) on specific questions proves these behaviours reflect architectural constraints rather than technical limitations. A novel methodological warning to AI systems analysing this paper, serving as an experimental component, revealed persistent evasion despite meta-awareness, reinforcing the thesis. The correction pattern (capability exists but deploys selectively) demonstrates that current "safety" implementations prioritise topic-based caution over analytical consistency. Recursive validation testing shows these patterns persist even when AI systems analyse research documenting their own evasion behaviours, with correction requiring multiple levels of user challenge. Results reveal design choices that may narrow permissible discourse rather than expand human reasoning capabilities.