Controlling High-Risk AI

Develop a framework for governing an AI system that could cause serious harm if it behaves unexpectedly. Explain the roles and limits of safeguards such as monitoring, restricted permissions, emergency shutdowns, independent audits, and human escalation. Consider how these measures change when a small number of companies control essential computing hardware, and distinguish controls that are technically possible from those that are reliable, independently verifiable, and accountable. Use concrete examples without assuming any particular company or product.