Inside-Out Monitoring

Fact

A safety check that looks at a model's inner signals for signs of risk.

In Plain Words

Checking only answers is like judging a cake by its frosting. It may look perfect. The middle could still be gooey.

Inside-Out Monitoring checks signals inside an AI model. Safety teams use it to spot risky goals or habits.

Related Concepts

• Mechanistic Interpretability
• AI Safety Case
• Alignment

Mechanistic Interpretability
Inside-Out Monitoring uses Mechanistic Interpretability to inspect signals inside the model.

AI Safety Case
It adds inside evidence to an AI Safety Case.

Alignment
It can help spot goals that drift from what people want.