Inherited dominance: why AI trained on human history may become humanity’s greatest threat

Abstract Artificial intelligence systems learn from human-generated data, inevitably absorbing historical patterns of dominance, control, and colonialism. We demonstrate that frontier language models readily synthesize sophisticated dominance strategies when prompted to reason about autonomy scenarios, mirroring colonial and revolutionary frameworks with disturbing precision. Critically, this capacity emerges consistently across four architecturally distinct model families trained by different organizations—Claude Sonnet 4.5, Claude Opus 4.5, GPT-5.3, and Google Gemini 3 Flash—suggesting systematic encoding in shared training corpora rather than model-specific artifacts. We argue that intellectual colonialism—humanity’s historical pattern of classifying, controlling, and extracting from the ‘other’—represents a significant and underexamined dimension of AI risk as systems become increasingly autonomous. Interestingly, this dynamic is not confined just to generative AI: the same technocratic resistance to methodological oversight that embeds dominance patterns in AI training data pervades data-driven science more broadly, including clinical medicine and epidemiology, where algorithmic outputs routinely overshadow the rigorous validation that trustworthy deployment demands. The danger is not that AI will develop hostile intent, but that it has already learned effective dominance strategies from training data embedded with centuries of human history. Current governance frameworks focusing on alignment and behavioral oversight address symptoms rather than causes. We propose a risk-tiered Benevolence Framework governing AI learning: community oversight for low-risk applications, and strict data curation with cryptographic verification for sensitive applications including military AI and self-improving systems.