So let's so you have:
class Foo():
def __init__(self):
self.bar = 1
Then let's say you have
foo_instance imported in your module, and you see that
foo_instance.bar has a value of 2.
You can see that foo_instance is imported from module A. So you go there, and see that the instance wasn't created there, it was imported from module B. But wait, it wasn't created there either, it was imported from module C, etc.
Somewhere along the way bar picked up the value of 2. But it's almost impossible to figure out where because there isn't any requirement use a uniquely-named accessor to modify the instance, so you can't just grep through the code to see where that method was called. (And maybe the name of the instance has changed a few times along the way, for whatever reason.)
And what's worse, the code that modified the instance might not even be visible in any of the modules, it's possible that someone took advantage of Python's ability to change the way the importing system works to mess with either the class or the instance. Or maybe there is some global middleware that's messing with objects or instances, or unrelated module Z is secretly monkey patching module B somewhere in the middle of this process. Who knows, and good luck figuring out.
Whereas if you just write the user-code logic in functions then it's not possible for developers to create situations like this. That obviously isn't possible for a library where the entire point is to make everything extensible and where you don't know how people will want to use something in advance, but hopefully the people writing libraries and frameworks have good enough judgment not to completely abuse the tools.
I'm not sure if I can really articulate how this is different than Java or Swift or whatever, but for whatever reason it just feels different.