# What is Causal Scrubbing? Definition & Technical Mechanics

> **Source:** [SPIDITS AI AI Glossary](https://spidits.com/ai-glossary)  
> **Category:** Mechanistic Interpretability  
> **Canonical Citation:** [https://spidits.com/ai-glossary/causal-scrubbing](https://spidits.com/ai-glossary/causal-scrubbing)

## Definition
Causal Scrubbing is an empirical evaluation framework designed to test mechanistic interpretability hypotheses. It works by replacing internal network activations with resampled activation distributions that preserve the hypothesis's theoretical invariants, measuring whether model performance degrades when the hypothesis is violated.

## Architectural Mechanics & Deep Dive
Causal Scrubbing is a quantitative framework introduced by Redwood Research for rigorously testing mechanistic interpretability hypotheses. Rather than relying on qualitative feature visualizations, Causal Scrubbing formalizes a hypothesis about a neural network circuit as a computational graph mapping. It then performs behavior-preserving resampling ablations, replacing internal activations with resampled equivalents that maintain the hypothesized invariants. If the model's output performance remains intact, the hypothesis is confirmed; if performance drops, the hypothesis is proven false.

## Industry Applications & Real-World Use Cases
AI safety audits, circuit hypothesis validation, and neural network interpretability verification

## Frequently Asked Questions

### What is Causal Scrubbing in AI research?
Causal Scrubbing is an automated algorithm for rigorously evaluating whether a proposed computational circuit hypothesis actually explains a neural network's behavior by systematically swapping activations.

### How does Causal Scrubbing differ from standard activation patching?
Unlike single-node activation patching, Causal Scrubbing converts complex circuit hypotheses into full trees of resampled distributions, testing entire structural invariants simultaneously.

## Related Terminology
- [explainable-ai](https://spidits.com/ai-glossary/explainable-ai)
- [ai-safety](https://spidits.com/ai-glossary/ai-safety)
- [neural-network](https://spidits.com/ai-glossary/neural-network)
- [deep-learning](https://spidits.com/ai-glossary/deep-learning)

---
*From the SPIDITS AI Knowledge Base (336+ terms indexed). Explore full technical definitions: [https://spidits.com/ai-glossary](https://spidits.com/ai-glossary)*