This would fix help fix #9008.

From most to least-well defined

1. Object arrays of float,np.float32, np.datetime64... → return np.isnan(item)
2. Object arrays of int,np.int32, ... → return False
3. Object arrays of any type that implements __float__ or __int__ - return np.isnan(float(x))
4. Object arrays of np.ndarray → return np.isnan(item) (requires a O->O loop, not O->?)
5. Any value → return x != x

I think I'd draw the line after 3, and have 4 onwards throw an error.