Demonstration showing a bug in Pandas: it automatically converts datetime columns into a different pandas-specific type, even when the original column has `dtype=object`.

bug_pandas_map_mangles_column_datatype.py

from datetime import datetime

import pandas as pd

now = datetime.now()

df = pd.DataFrame.from_dict(
    {
        "created_at": pd.Series([now, now - timedelta(seconds=100), now + timedelta(seconds=10)], dtype='object'),
    }
)

# this is True
assert df['created_at'].map(lambda x: isinstance(x, datetime)).all()

identity = df['created_at'].map(lambda x: x)

# THIS WILL FAIL!
# Even though we are doing _nothing_ to the columns values!
assert identity.map(lambda x: isinstance(x, datetime)).all()
# Pandas as silently changed the type of the column from datetime to datetime64[ns] :(

Behavior observed in pandas==1.5.3