Correlation is Not Causation: An Intro to Causal AI with pgmpy

Every ML model you've trained does one thing: find patterns. Given enough data, a model will confidently tell you that ice cream sales predict drowning rates. The correlation is real. The conclusion is wrong. And the model has no way to know the difference.

This isn't a toy problem. It's the precise failure mode behind production ML systems that make decisions about people - hiring pipelines that penalise candidates through proxy variables, medical models that confuse "patients who received this drug" with "patients who would benefit from it", recommendation systems optimising for proxies of what users actually want. The model learned the data correctly. It simply cannot answer the question you're actually asking, because you're asking a causal question and it was only trained to answer a statistical one.

The gap between those two questions has a name, and a rigorous framework behind it.

In this talk, I'll walk you through why correlation-based models fail structurally for decision-making not vaguely, but precisely, using Pearl's Ladder of Causation as the framework. Seeing, doing, and imagining: three levels of causal reasoning that standard ML never climbs past the first of. From there we move into causal graphs how to read a DAG, what directed and bidirected edges encode, and what the do-operator actually means when you stop treating it as notation and start treating it as a question. The talk closes with a live demo in pgmpy: building a causal graph, running a causal identification query, and getting an answer that a regression model structurally cannot give you.

The goal is not to convince you to throw away your existing models. It's to give you a precise vocabulary for a class of questions your models can't answer, and a practical tool for when you need to answer them.

• Why ML models fail silently for decision-making and the structural reason behind it

• The intuition behind causal graphs, DAGs, and what "intervention" actually means in practice

• A hands-on first look at pgmpy's causal identification module with runnable code

• How to start thinking causally about problems you're already solving with ML