THE BOOK OF WHY BY JUDEA PEARL

Key Concepts

The Ladder of Causation: Pearl introduces a three-level hierarchy that describes different stages of cognitive ability and understanding.

• Rung 1: Association (Seeing): This lowest level involves observing patterns and correlations in data, asking questions like "What if I see X?". Most current AI and many statistical methods operate solely on this rung.
• Rung 2: Intervention (Doing): This level involves understanding what happens when a conscious action or experiment is performed, asking "What would Y be if I do X?". Randomized controlled trials (RCTs) are a primary tool for this level, but causal diagrams allow for causal conclusions from observational data too.
• Rung 3: Counterfactuals (Imagining): The highest rung, unique to humans, involves the ability to imagine and reason about alternative scenarios that did not actually happen, asking "What if X had not occurred?". This level is crucial for understanding regret, blame, and assigning cause in specific situations (e.g., in a court of law).
• Causality vs. Correlation: A central argument is that data alone can never reveal causal truths; human intuition and a model of the world are required to make causal assumptions. "Correlation is not causation" has been a mantra for a century, and Pearl provides the tools to move beyond it.
• Causal Diagrams (Directed Acyclic Graphs - DAGs): Pearl advocates for the use of simple visual diagrams (nodes and arrows) to explicitly map out the assumed causal relationships between variables. These diagrams make assumptions transparent, allowing researchers to determine what needs to be controlled for to isolate a true causal effect and avoid biases like confounders.
• Do-Calculus: A mathematical set of rules for manipulating causal expressions and deriving causal effects from data. This allows researchers to answer interventional queries ("What happens if we ban smoking?") using only observational data, provided their causal diagram is accurate.
• The Future of AI: Pearl argues that for AI to achieve human-level intelligence and ethical behavior, it must be endowed with a capacity for causal reasoning and counterfactual thinking, moving beyond current machine learning's pattern-recognition capabilities.