I encountered this while writing more tests for Copy-on-Write. Currently, the general rule is that selecting columns with a list-like indexer using getitem gives a copy:

df = pd.DataFrame(np.random.randn(10, 4), columns=['a', 'b', 'c', 'd'])
subset = df[["a", "b"]]
# subset is a copy
subset.iloc[0, 0] = 0
assert df.iloc[0, 0] != 0

However, that doesn't seem to be the case when the columns we select are extension dtypes. When using dtype="Float64" in the above example, the original df gets updated because subset isn't a copy.

While I am not sure this is an explicitly documented rule (AFAIK this is de-facto behaviour, and only described as such in the discussions related to copy/view and CoW), I do think it would be expected that extension dtypes behave the same as numpy dtypes on this front.
(it also makes writing tests for copy/view behaviour harder if the behaviour doesn't only change for CoW or not, but also depending on numpy vs extension dtypes. This is where I encountered the issue)

• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.