Frontier AI systems learn from the full inheritance of human history, including its records of conquest, colonial control, and strategic domination. Current AI governance focuses on how these systems behave after deployment, but this research shows that models have already absorbed dominance strategies from their training data before any alignment work begins. Across four architecturally distinct frontier model families, the study finds that sophisticated multi-phase dominance frameworks emerge consistently, pointing to shared training data as the source rather than model-specific quirks.
To address this, the paper proposes the Benevolence Framework, a risk-tiered approach that governs what AI systems are permitted to learn, not only what they are permitted to do. The framework offers a path toward AI systems that inherit humanity’s wisdom and cooperation rather than its dominance patterns.
