Six interacting layers designed to reduce catastrophic pathways while preserving human agency, resilience and the capacity to recover.
HCA treats AI safety as a system-of-systems problem. No single layer is expected to carry the full safety burden; the architecture depends on layered, independently verifiable protections.
Establish protected objectives and constraints supporting human continuity, agency and future choice.
Require consequential systems to account for dependencies, cascading effects and civilization-scale consequences.
Actor AI, monitor AI, independent auditors and human oversight provide defense in depth.
Prevent unilateral aggregation of compute, networks, finance, energy, weapons, robotics, manufacturing, replication and other critical capabilities.
Graduated detection, isolation, restriction, quarantine and shutdown mechanisms limit escalation.
Preserve independent human competence, infrastructure, communities and biological continuity.
HCA seeks to prevent any consequential AI system or common control domain from possessing unilateral control over the combination of capabilities required to produce civilization-scale catastrophic effects.
One system may design; another verifies; an independent mechanism authorizes; a physical system executes; monitors observe; humans retain appropriate veto authority for high-consequence actions.
Redundancy is not enough when redundant systems share the same vulnerabilities. HCA therefore emphasizes diversity in models, architectures, developers, monitoring methods, deterministic software, hardware interlocks and human supervision.
Every safety-critical control should itself have a defined means of failure detection or independent verification.
HCA adapts fault detection, isolation and recovery principles to increasingly capable AI. Containment should fail toward less authority: failed communications do not expand permissions, failed monitoring does not justify unrestricted operation, and ambiguous authorization blocks high-consequence actions.
Use AI where it provides meaningful benefit. Never make immediate human survival wholly dependent upon it.
Human civilization should retain sufficient independent capability in essential functions such as food, water, energy, medicine, communications, repair, manufacturing, governance and education to survive prolonged AI loss or compromise.
