diff --git a/xarray/core/variable.py b/xarray/core/variable.py
index 1ea86254d35..52088090468 100644
--- a/xarray/core/variable.py
+++ b/xarray/core/variable.py
@@ -210,8 +210,9 @@ def _maybe_wrap_data(data):
 
 def _possibly_convert_objects(values):
     """Convert object arrays into datetime64 and timedelta64 according
-    to the pandas convention.  For backwards compat, as of 3.0.0 pandas,
-    object dtype inputs are cast to strings by `pandas.Series`
+    to the pandas convention. Object dtype inputs that are inferred to be
+    strings are returned unchanged. For backwards compat, as of 3.0.0 pandas,
+    the remaining object dtype inputs are cast to strings by `pandas.Series`
     but we output them as object dtype with the input metadata preserved as well.
 
 
@@ -220,8 +221,22 @@ def _possibly_convert_objects(values):
     * pd.Timestamp
     * pd.Timedelta
     """
-    as_series = pd.Series(values.ravel(), copy=False)
-    result = np.asarray(as_series).reshape(values.shape)
+    inferred = pd.api.types.infer_dtype(values.ravel(), skipna=True)
+
+    if inferred == "string":
+        return values
+    elif inferred == "datetime":
+        result = pd.to_datetime(values.ravel()).to_numpy().reshape(values.shape)
+    elif inferred == "timedelta":
+        result = pd.to_timedelta(values.ravel()).to_numpy().reshape(values.shape)
+    elif inferred in ["datetime64", "timedelta64"]:
+        # Casting drops unit info for these cases;
+        # fall back to pd.Series roundtrip, which preserves them.
+        as_series = pd.Series(values.ravel(), copy=False)
+        result = np.asarray(as_series).reshape(values.shape)
+    else:
+        return values
+
     if not result.flags.writeable:
         # GH8843, pandas copy-on-write mode creates read-only arrays by default
         try:
diff --git a/xarray/tests/test_variable.py b/xarray/tests/test_variable.py
index c2bc73f70b9..0ebde435697 100644
--- a/xarray/tests/test_variable.py
+++ b/xarray/tests/test_variable.py
@@ -70,6 +70,9 @@ def var():
     [
         np.array(["a", "bc", "def"], dtype=object),
         np.array(["2019-01-01", "2019-01-02", "2019-01-03"], dtype="datetime64[ns]"),
+        np.array([datetime(2000, 1, 1), datetime(2000, 1, 2)], dtype=object),
+        np.array([timedelta(seconds=1), timedelta(seconds=2)], dtype=object),
+        np.array([1, "a"], dtype=object),
     ],
 )
 def test_as_compatible_data_writeable(data):