Causal Inference, Counterfactuals, and Decision AI
Beyond Correlation
Why causality matters
Predictive accuracy alone is insufficient when interventions change the data-generating process. Causal AI seeks models that support reasoning about actions, policy changes, transportability, and counterfactual outcomes.
Causal vocabulary
Intervention notation
Causal effects are often expressed with -calculus and structural causal models. For example, the causal effect of on is generally not equal to the observational association .
Causal questions
| Question type | Example | Why it is hard |
|---|---|---|
| Prediction | Will this user click? | May be answered by correlation |
| Intervention | What if we change the ranking policy? | Requires modeling policy shift |
| Counterfactual | Would the user have clicked under a different ranking? | Requires unit-level latent reconstruction |
| Transportability | Will the model work in another population? | Requires understanding domain shift |
Which concept distinguishes causal reasoning from ordinary prediction?
Causal reasoning asks what happens if we intervene, not merely what is associated in the observed data.
Correct answer: Interventions
Why can a model with excellent observational accuracy still fail under intervention?
Causal structure is needed for stable decision-making across interventions.
Correct answer: Because it may exploit spurious correlations that break when the environment or policy changes.
Open research direction
Causal representation learning aims to infer latent causal variables from high-dimensional observations, but identifiability and evaluation remain difficult.
Approaching a causal AI problem
-
1
Step 1: Define the causal question precisely.
-
2
Step 2: Specify assumptions with a structural causal model or graph.
-
3
Step 3: Identify confounders, mediators, and colliders.
-
4
Step 4: Derive an estimand and choose an estimator.
-
5
Step 5: Stress-test robustness with sensitivity analysis.