Feature Importance — MDI Limitations

You train a RandomForestClassifier to predict customer churn and extract feature importances. customer_id (a random integer unique per customer) appears as the third most important feature.

1. Why is customer_id appearing as important?
2. What is the fundamental flaw in MDI (Mean Decrease in Impurity) that causes this?
3. How would you get reliable feature importances for this problem?