Mechanistic interpretability is an AI research field that reverse-engineers trained neural networks to explain, in human-understandable terms, which features a model represents and how those features combine through circuits to produce its outputs. Unlike approaches that look only at the relationship between inputs and outputs, it analyzes a model's internal activations and weights directly.
It has drawn attention as a way to examine the safety and alignment of large language models. A common experimental method is to locate an internal representation and increase or decrease it to observe changes in behavior.