A mechanic who wants to know why a car stalls doesn’t stop at watching it drive. They open the hood and trace which part does what. Mechanistic interpretability takes that approach to a neural network. Instead of only testing what a model says, researchers look at its internal activations and try to identify the pieces of computation behind an answer.
Two ideas organize the field. A feature is a direction in the network’s internal activity that corresponds to a concept, such as a city, a programming syntax error or a famous person. A circuit is a set of connected features that performs a step, the way early work on image models traced curve detectors combining into a circle detector. The hard part is that single neurons usually mix many concepts, a problem researchers attribute partly to superposition: the network stores more features than it has neurons. Anthropic’s work uses sparse autoencoders, a separate model trained to pull those mixed signals apart into cleaner features, and in 2024 reported millions of features in Claude 3 Sonnet. Turning some of them up or down changed the model’s behavior.
The field is a narrower, more causal relative of explainability. Explainability tools usually describe which inputs mattered for an output; mechanistic interpretability tries to show how the model computed it. Researchers hope it will help detect hidden failures such as scheming that behavioral tests can miss, and it is a major strand of AI safety research. It remains early: current methods explain only part of a large model’s behavior, and the field has not yet produced a full explanation of how a frontier model answers.
Sources
- Bereska and Gavves: Mechanistic Interpretability for AI Safety: A Review (2024) — Defines the field as reverse engineering the mechanisms and representations learned by neural networks into human-understandable algorithms and concepts, and surveys its methods; checked October 24, 2026.
- Olah et al.: Zoom In: An Introduction to Circuits (Distill, 2020) — Shows circuits inside an image model, such as curve detectors combining into a circle detector, and argues networks can be studied at this scale; checked October 24, 2026.
- Bricken et al.: Towards Monosemanticity: Decomposing Language Models With Dictionary Learning (Anthropic, 2023) — Uses a sparse autoencoder to extract interpretable features from a one-layer transformer and discusses superposition, where a network represents more features than it has neurons; checked October 24, 2026.
- Templeton et al.: Scaling Monosemanticity (Anthropic, 2024) — Extracts features from Claude 3 Sonnet, Anthropic's production model at the time, including abstract and safety-relevant features that both respond to and can steer behavior; checked October 24, 2026.