Consider the following code sample :

keys = ['x', 'y', 'z']
d = dict.fromkeys(keys, [])
d['x'].append('dont')
d['y'].append('mix')
d['z'].append('me!')
print d['x']

>>> ['dont', 'mix', 'me!']

It is very unatural and dangerous to have all dict keys poining to the
same mutable object reference. 

The way it should behave : 
if value is mutable, create a new copy of value for each keys
else, it doesn't matter

This behavior will never change. Use collections.defaultdict for your case.

I concur.

It does not suits my needs.

what if keys are passed as an argument?

I have to add this ugly loop :

for key in keys:
    d[key] = [] 
    

I really think that :
1) The doc should warn about it, since it is a very weird behaviour
2) There could at least be a third argument allowing the user to choose,
with a default value that maintains the current behaviour (for
compatibility, even though I think the default should be a deep copy)

The docs very clearly say the second argument is the value that will be
assigned to the keys.  It doesn't matter whether or not that object is
mutable, it is that object that gets assigned.  This is just the way
that Python works:

>>> a = b = []
>>> a.append(1)
>>> b.append(2)
>>> a
[1, 2]

That is exactly analogous to what you are doing in your example.

I'll add a bit of explanation as well:

> I have to add this ugly loop :
> 
> for key in keys:
>    d[key] = [] 
 
With a defaultdict, you don't -- you just use d[key], and if not already
present, the entry will be initialized with an empty list.

> I really think that :
> 1) The doc should warn about it, since it is a very weird behaviour

It is not weird in Python.  Implicit copies are never made, partly
because it's very hard to implement it correctly for every object.

> 2) There could at least be a third argument allowing the user to choose,
> with a default value that maintains the current behaviour (for
> compatibility, even though I think the default should be a deep copy)

The interface of a fundamental type like "dict" will not be changed
lightly. Usually, when a change is made, it is to add a feature that is
often requested by lots of users -- I've never seen someone request this
before.  And as we've seen, it can easily be written either using a
simple loop, or a defaultdict, depending on the exact use-case.