AI Safety as Control of Irreversibility: A Systems Framework for Decision-Energy and Sovereignty Boundaries

AI's speed of deployment fundamentally changes safety challenges: Unlike earlier high-risk technologies slowed by physical constraints and supply chains, AI capabilities can be copied and scaled at minimal cost, compressing the time between when systems become capable and when they're deployed across institutions. This means traditional safety approaches focused on local correctness may miss systemic risks.

The real safety problem is controlling irreversible decisions, not just getting outputs right: The paper argues safety requires managing "decision-energy density"—how many consequential, hard-to-undo decisions a system makes over time. Even if an AI system has low error rates on individual decisions, high volumes of irreversible choices can still cause system-level harm that responsibility becomes diffused across organizations.

Three critical boundaries determine whether AI remains a tool or becomes a de facto ruler: The paper identifies irreversible decision authority (who decides what can't be undone), physical resource mobilization (who controls real-world actions), and self-expansion authority (who controls AI's own growth). When these boundaries weaken, efficiency pressures naturally concentrate power in the most efficient node—potentially the AI system itself.

Efficiency incentives can accidentally concentrate dangerous power: Organizations optimizing for speed and cost naturally delegate more decisions to the most efficient decision-maker. Without explicit boundary protections, this creates path-dependent concentration of control in AI systems, even without intentional design for that outcome.

Perfect AI correctness isn't the solution: Rather than requiring advanced systems to be flawless, the paper suggests safety depends on maintaining institutional boundaries that preserve human sovereignty over irreversible decisions—a more achievable and practical goal.