My colleague, Jim Forrester, has his own paradox (Forrester, J. 1984.Gentle murder, or the adverbial samaritan, Journal of Philosophy, 81: 193-197. If you have subscription, you can view it at JStor.). Lucky guy.
Each of the following seem to be quite reasonable deontic principles:
(i) [O(p --> q) --> .p --> Oq] (where O is a deontic operator meaning "it is obligatory that")
(ii) (p --> q) --> (Op --> Oq) (inference rule)
(iii) Op --> ~O~p
Unfortunately, they seem to lead to paradox. For suppose that we have a legal system which, while prohibiting all forms of murder, nevertheless regards brutal murders to be less wrong than gentle murders. On the supposition that there is such a legal system, we get the following argument:
(1) O~(Smith murder Jones) [assumption]
(2) O(Smith murders Jones --> Smith murder Jones gently) [assumption]
(3) Smith murders Jones --> O(Smith murder Jones gently) [(i) & (2)]
(4) Smith murders Jones [assumption]
(5) O(Smith murder Jones gently) [(3) & (4)]
(6) Smith murders Jones gently --> Smith murders Jones [logic]
(7) O(Smith murder Jones gently) --> O(Smith murders Jones) [(ii) & (6)]
(8) O(Smith murders Jones) [(5) & (7)]
(9) ~O~(Smith murders Jones) [(iii) & (8)]
And now we have our contradiction. As Jim notes, (i) seems to be the questionable deontic principle--though he thinks it holds up well enough in the case as described.
I'm not so sure. Although I do think that something like (i) is pretty widely accepted, I'm not sure it has the right formal structure to push the paradox through. Specifically, I'm not sure that the wide scope reading of O in the antecedent of (i) is ever acceptable.
Let me change the example (because I find it clearer). In the animal welfare debate, one sometimes runs across people who have become convinced that meat-eating is wrong but who still persist in their carnivorous ways. In an attempt to make themselves less morally culpable, they often say things such as "Well, at least they kill them quickly." The idea being that, even if it is wrong to kill animals just because they are yummy, if we are going to do it anyway we ought to do it quickly.
Now in this case, I find it extremely difficult to find a plausible representation of this reasoning which would involve the conditional O(kill animals --> kill animals quickly). To my ear, this is categorially wrong. What does it mean to say that a conditional is obligatory? Acts are the sorts of thing that are obligatory or not, and conditionals don't offer up any acts in the right sorts of "positions". It seems, rather, that what is obligatory is that conditions be made such that if we kill animals, then we kill them quickly. But this can't be reasonably represented as an instance of O(p --> q); it is just an instance of Op.
[Addition: the oddness I am gesturing at is also reflected in the oddness of formalizing "It is obligatory that Smith not murder Jones" as (1). (1) translates as "it is obligatory that it not be the case that Smith murder Jones," which I find horrendous. It seems like there must be some Davidsonion analysis that makes all of this perspicuous.]
We can represent the latter proposition a bit more explicitly, if you like, by making notationally clear what the conditional circumstances are. Thus: O(condition[kill animals --> kill animals quickly]). But I don't think it follows from this plus the fact that we kill animals, that we are obligated to kill them quickly. That is, I deny (i*):
(i*) [O(condition[p --> q]) --> .p --> Oq]
Rather, it seems to me that the principle is acceptable only if Op; that is, only if:
(i**) [O(condition[p --> q]) --> .Op --> Oq]
But then, clearly the argument won't go through.