Mechanistic Interpretability
The subfield of AI safety concerned with reverse-engineering neural networks — understanding what individual neurons, attention heads, and circuits actually compute. Sparse autoencoders, attribution graphs, and probing are common techniques.
Related terms
- Interpretability
The subfield concerned with understanding what happens inside AI models — which features they represent, how they compose, why they fire on given inputs. Mechanistic interpretability is a leading research direction; sparse autoencoders are one recent technique.
- Sparse Autoencoder (SAE)
A neural network trained to reconstruct model activations from a sparse dictionary — the leading interpretability technique of 2024-2026. Enables mechanistic study of what individual features inside frontier LLMs represent. Anthropic and OpenAI have both published SAE-based interpretability work.
- Alignment
The general problem of making an AI system behave according to its designers' or operators' intent, especially as capability grows. Practical alignment work spans RLHF, Constitutional AI, red-teaming, evaluation, and interpretability research.
Related on this site
Where this fits
Mechanistic Interpretability is part of the Evaluation & Governance vocabulary used in the Generative AI Maturity Framework. See the full glossary for the complete set of 149 defined terms, or take the free maturity assessment to see where your organisation stands.