## TL;DR
Causal AI moves beyond correlation-based prediction to reason about cause and effect — enabling models to answer "what if" questions, make robust decisions under distribution shift, and avoid spurious correlations that break ML systems in production.

## Core Explanation
Correlation ≠ causation: an AI that learns "umbrella sales predict rain" fails when umbrella sellers run promotions during dry weather. Causal models explicitly represent the data-generating process via Structural Causal Models (SCMs): directed acyclic graphs where nodes are variables and edges represent causal relationships. do-calculus provides rules for transforming expressions with do(X=x) interventions into estimable quantities.

## Detailed Analysis
Key frameworks: (1) Potential Outcomes (Rubin) — each unit has multiple potential outcomes under different treatments; (2) SCM (Pearl) — equations encode mechanisms, interventions break incoming arrows; (3) Structural Equation Modeling. Modern ML applications: causal representation learning (locate independent causal mechanisms), causal fairness (remove discriminatory causal paths), and causal RL (learn interventions that transfer across environments).

## Further Reading
- CausalML Book (causalml-book.org)
- DoWhy + EconML Libraries
- Stanford Causal AI Lab