Which term signifies the field that explores how to interpret the inner workings of neural networks by examining their constituent parts?

• A. Model Organisms of Misalignment
• B. AI Welfare
• C. Mechanistic Interpretability
• D. AI Safety

Answer: (C)

Mechanistic interpretability studies how to interpret the inner workings of neural networks by examining their constituent parts—neurons, layers, and modules—and tracing how these pieces combine to produce a model’s behavior. This field aims to map specific components and circuits to the computations they perform, often by analyzing activations, probing how information flows, and reconstructing small mechanistic pieces that drive decisions. It’s the best fit here because it directly targets understanding what parts of the model are doing and how they interact to yield outcomes, rather than addressing broad safety concerns or welfare considerations. AI Safety is a broader umbrella about preventing harm and ensuring reliable behavior, not necessarily dissecting internal mechanisms. AI Welfare focuses on wellbeing-related questions, not the technical decoding of internal neural processes. The term Model Organisms of Misalignment isn’t an established field describing interpretability work.