The Constraint Fallacy

There is a popular approach to AIArtificial IntelligenceSystems that learn, adapt, and act with real-world impact governance that we need to address directly: constraining the tool.

Guardrails. Output filters. Prompt injection defenses. Rate limiting. Sandboxing. These are all forms of constraining the instrument — limiting what the tool can do.

This is analogous to governing a human by constraining their hands. You can put gloves on someone to prevent them from picking locks. You can put them in a straitjacket to prevent them from moving. But you haven't governed the person. You've restrained an appendage.

The agent — the entity making decisions — is upstream of the tool. Constraining the tool doesn't make the agent accountable. It makes the tool less useful while leaving the decision-making process ungoverned.

In practice, this is why AIArtificial IntelligenceSystems that learn, adapt, and act with real-world impactsafety efforts focused purely on model behavior keep getting circumvented. The model is the hand, not the brain. Jailbreaks work because they target the decision-making layer that the constraint layer can't see. Output filters work until someone figures out how to ask the question differently.

Real governance addresses the decision-making entity, not the instrument it uses. It provides awareness, incentive, and consequence at the level where decisions are actually made.