Information Geometry of AI Deployment: Why Architecture Determines Safety Outcomes

Key Points

• The aim is to explore how deployment geometry impacts AI safety more than model alignment.
• Constructed the deployment manifold using Bernoulli statistical manifolds.
• Applied Čencov-unique Fisher-Rao metric for analysis.
• Formulated six open problems linking information geometry to AI safety.
• Proved that deployment geometry is a stronger determinant of AI safety outcomes than model alignment.
• Identified six critical research questions for future exploration.

Abstract