diff --git a/_questions/machine-learning-zoomcamp/module-3/033_08b4b5de10_pd-get-dummies-vs-dictvectorizer-one-hot-q4.md b/_questions/machine-learning-zoomcamp/module-3/033_08b4b5de10_pd-get-dummies-vs-dictvectorizer-one-hot-q4.md
new file mode 100644
index 00000000..57d89825
--- /dev/null
+++ b/_questions/machine-learning-zoomcamp/module-3/033_08b4b5de10_pd-get-dummies-vs-dictvectorizer-one-hot-q4.md
@@ -0,0 +1,16 @@
+---
+id: 08b4b5de10
+question: Can I use `pd.get_dummies` instead of `DictVectorizer` for one-hot encoding
+  (Homework 3 Q4)?
+sort_order: 33
+---
+
+Yes, you can use `pd.get_dummies`, but `DictVectorizer` is safer and is what the lectures use.
+
+The main risk with `pd.get_dummies` is that it is applied separately to each split. If a category appears in training but not in validation (or vice versa), the encoded feature matrices can end up with a different set of columns. That can either crash your pipeline or, worse, silently misalign features.
+
+With `DictVectorizer`, you fit on training only and then transform validation/test with the same learned mapping, so the column layout is locked. Unseen categories at transform time become all-zero columns, which is the intended behavior.
+
+If you still use `get_dummies`, make sure you explicitly align the columns between splits, for example by reindexing validation to match training’s columns and filling missing values with 0.
+
+Also, when using `DictVectorizer`, note that `get_feature_names_out()` lets you inspect the resulting feature/column names.
\ No newline at end of file