# Interpretability and Transparency

Interpretability aims to understand and explain what AI models do internally. This sub-article expands [Section 4.2 of Philosophy](../philosophy.md).

## Why Interpretability Matters
- Trust, safety, debugging, and accountability.
- Without it, we cannot reliably predict failures or correct misalignment.

## Approaches
- **Mechanistic interpretability**: reverse-engineering internal representations (e.g., sparse autoencoders, feature circuits).
- **Post-hoc explanations**: saliency maps, attention patterns, and natural-language justifications.
- **Transparency**: publishing system cards, data cards, and evaluation results.

## Limitations
- Explanations may be unfaithful to the actual computation.
- Large models resist simple global explanations.
- Trade-offs with performance and privacy.

## Why It Matters
Interpretability is a practical tool for alignment and governance, not just a philosophical goal.

## Sources
- Olah, C. et al. (2020). *Interpretability via feature extraction.*
- Anthropic. *Sparse autoencoders research.*

*Back to [Philosophy](../philosophy.md).*