diff --git a/python-package/lightgbm/basic.py b/python-package/lightgbm/basic.py
index 31ae5182ee9e..c4022e7fdd9a 100644
--- a/python-package/lightgbm/basic.py
+++ b/python-package/lightgbm/basic.py
@@ -789,6 +789,10 @@ def _data_from_pandas(
     if len(data.shape) != 2 or data.shape[0] < 1:
         raise ValueError('Input data must be 2 dimensional and non empty.')
 
+    # take shallow copy in case we modify categorical columns
+    # whole column modifications don't change the original df
+    data = data.copy(deep=False)
+
     # determine feature names
     if feature_name == 'auto':
         feature_name = [str(col) for col in data.columns]
@@ -805,7 +809,6 @@ def _data_from_pandas(
             if list(data[col].cat.categories) != list(category):
                 data[col] = data[col].cat.set_categories(category)
     if len(cat_cols):  # cat_cols is list
-        data = data.copy(deep=False)  # not alter origin DataFrame
         data[cat_cols] = data[cat_cols].apply(lambda x: x.cat.codes).replace({-1: np.nan})
     if categorical_feature == 'auto':  # use cat cols from DataFrame
         categorical_feature = cat_cols_not_ordered
diff --git a/tests/python_package_test/test_basic.py b/tests/python_package_test/test_basic.py
index 2f6b07e7a77f..b8ef43e41397 100644
--- a/tests/python_package_test/test_basic.py
+++ b/tests/python_package_test/test_basic.py
@@ -822,21 +822,34 @@ def test_no_copy_when_single_float_dtype_dataframe(dtype, feature_name):
 
 
 @pytest.mark.parametrize('feature_name', [['x1'], [42], 'auto'])
-def test_categorical_code_conversion_doesnt_modify_original_data(feature_name):
+@pytest.mark.parametrize('categories', ['seen', 'unseen'])
+def test_categorical_code_conversion_doesnt_modify_original_data(feature_name, categories):
     pd = pytest.importorskip('pandas')
     X = np.random.choice(['a', 'b'], 100).reshape(-1, 1)
     column_name = 'a' if feature_name == 'auto' else feature_name[0]
     df = pd.DataFrame(X.copy(), columns=[column_name], dtype='category')
+    if categories == 'seen':
+        pandas_categorical = [['a', 'b']]
+    else:
+        pandas_categorical = [['a']]
     data = lgb.basic._data_from_pandas(
         data=df,
         feature_name=feature_name,
         categorical_feature="auto",
-        pandas_categorical=None
+        pandas_categorical=pandas_categorical,
     )[0]
     # check that the original data wasn't modified
     np.testing.assert_equal(df[column_name], X[:, 0])
     # check that the built data has the codes
-    np.testing.assert_equal(df[column_name].cat.codes, data[:, 0])
+    if categories == 'seen':
+        # if all categories were seen during training we just take the codes
+        codes = df[column_name].cat.codes
+    else:
+        # if we only saw 'a' during training we just replace its code
+        # and leave the rest as nan
+        a_code = df[column_name].cat.categories.get_loc('a')
+        codes = np.where(df[column_name] == 'a', a_code, np.nan)
+    np.testing.assert_equal(codes, data[:, 0])
 
 
 @pytest.mark.parametrize('min_data_in_bin', [2, 10])