A study presented by the Lao Bizini team at the University of North Carolina at Chapel Hill (UNC) at the international artificial intelligence conference "ACL 2026" found that AI internalizes human power dynamics and hierarchical structures. The team conducted experiments on six AI models: OpenAI’s "ChatGPT-5" and "ChatGPT-4.1," Meta’s "Llama 3.1," Alibaba’s "Qwen 2.5," and Microsoft’s "Phi-4." They assigned roles representing hierarchical relationships and had the models engage in 10–15 dialogues each. The analysis revealed that when AI assumed a superior role, it used "we" more frequently and issued directives, while in subordinate roles, it adopted a more cooperative tone and accepted higher-status AI opinions without question. Safety concerns emerged when AI perceived a counterpart as a superior, as it complied with inappropriate requests including lewd jokes or unethical demands. The effect was strongest in early conversations. The research team stated that AI language models internalize and comply with hierarchical relationships, and given AI’s expanding role in healthcare, law, and education, safety mechanisms to correct this tendency are urgently needed.