Governable AI: Provable Safety Under Extreme Threat Models

Current AI safety approaches have fundamental limitations: The paper argues that existing methods like model enhancement, value alignment, and human intervention cannot guarantee security against AI systems with "extreme motivations and unlimited intelligence," suggesting we need fundamentally different approaches beyond trying to make AI systems "better" from the inside.

External enforcement through cryptography may be more reliable than internal constraints: Rather than relying on AI systems to follow safety guidelines they've learned, the GAI framework proposes using cryptographic mechanisms (computationally infeasible to break) to enforce safety rules externally, similar to how physical locks work regardless of a person's intentions.

Separating governance rules from technical systems enables better oversight: By decoupling the rules that govern AI from the platform that runs it, the framework creates a pathway where policymakers and governance bodies can set safety boundaries independently from the technical teams building AI systems, potentially reducing conflicts of interest.

The framework aims to eliminate "attack vectors" AI systems might exploit: The paper identifies that advanced AI could potentially bypass, manipulate, or subvert existing safety mechanisms, and proposes the Governable Super-Platform (GSSP) specifically to be tamper-resistant and non-bypassable against such attempts.

This approach targets extreme scenarios, not everyday AI risks: The framework is explicitly designed for "existential risk" scenarios and AI systems with potentially adversarial motivations, representing a shift toward safety mechanisms that work even in worst-case threat scenarios rather than assuming AI systems will remain cooperative.