Which term refers to research that tries to understand what is happening inside a neural network at the level of individual components?

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

Answer: (B)

Mechanistic interpretability is the study that aims to understand exactly what each part of a neural network is doing, down to individual components like neurons, attention heads, or small circuits. It involves mapping activations to functions, testing causal roles by perturbing parts, and building explanations of how information flows inside the model. The goal is to translate the model’s internal machinery into human-understandable mechanisms, so we can see why it produces certain outputs and anticipate failures or misbehaviors. This focus on the inner workings at the component level is what sets it apart from broader interpretability or safety efforts that look at behavior from the outside. The other terms don’t describe this inner-component analysis: one is not a standard term for internal circuitry, another centers on AI welfare, and another refers to a class of very large models.