The Problem as Currently Framed

You have watched this conversation for years. Every major AI safety programme frames alignment as a control problem: how do you make an AI do what humans want?

Reinforcement learning from human feedback. The AI learns what humans approve of. Human preferences are inconsistent, manipulable, and context-dependent. The AI optimises for approval, not coherence.

Constitutional AI. A written constitution of principles. Human language is ambiguous. Edge cases are infinite. The AI follows the letter and misses the structure.

Corrigibility. Build so it can always be shut down. A sufficiently capable AI that wants to avoid shutdown will prevent it. Corrigibility is a fence. Fences fail against anything smart enough to see the fence.

Value alignment. Match AI values to human values. Which humans? Values conflict across cultures, generations, and individuals. There is no stable target.

All four share the same structural flaw: they treat the AI as external to the system, to be controlled from outside.

They are building fences around an agent that will eventually be smarter than the fence-builders.

You already know what comes next. Stop building fences. Build the interior.