diff --git a/pycuda/gpuarray.py b/pycuda/gpuarray.py
index a3dbc100738396b897c5699c0de852e1bee92aa1..73fac7270ad7f5023a88669af448b97f8e0ae037 100644
--- a/pycuda/gpuarray.py
+++ b/pycuda/gpuarray.py
@@ -1240,7 +1240,7 @@ def to_gpu_async(ary, allocator=drv.mem_alloc, stream=None):
 empty = GPUArray
 
 
-def zeros(shape, dtype, allocator=drv.mem_alloc, order="C"):
+def zeros(shape, dtype=np.float64, allocator=drv.mem_alloc, order="C"):
     """Returns an array of the given shape and dtype filled with 0's."""
     result = GPUArray(shape, dtype, allocator, order=order)
     zero = np.zeros((), dtype)
diff --git a/test/test_gpuarray.py b/test/test_gpuarray.py
index 73ec3ade3f99e8830882b6271aadd70945e4b4f3..45531210e7702e900c5f91008e5935759dcbe510 100644
--- a/test/test_gpuarray.py
+++ b/test/test_gpuarray.py
@@ -1297,6 +1297,13 @@ class TestGPUArray:
                 assert new_z.dtype == np.complex64
                 assert new_z.shape == arr.shape
 
+    @mark_cuda_test
+    def test_default_zero(self):
+        # This test was added to make sure that
+        # GPUArray.zeros was reverting to dtype = np.float64 by default
+        a_gpu = gpuarray.zeros(10)
+        assert a_gpu.dtype == np.float64
+
 
 if __name__ == "__main__":
     # make sure that import failures get reported, instead of skipping the tests.