I think you should always cite a paper on the arXiv, if it would be appropriate to cite the paper at all.

I want to strongly encourage people to put preprints of the arXiv. It makes mathematics more open. It's incredibly useful for following what's going on in fields not quite your own, or if you are not in personal contact with everyone in your field. However, I know mathematicians who do not put preprints on the arxiv (or only do so just before publication) because they are worried about "losing citations". Sadly it seems that many funding agencies and universities are starting to use citation data; so it will become more important for citations not to be "lost".

Thus, it would seem that a culture of not citing arXiv papers (just because they are on the arXiv) is going to encourage people not to use the arXiv, which I think is a bad thing.

Of course, as Willie says, you should be very careful about proofs. But, as other questions on MathOverflow have suggested, just because a paper is formally published doesn't ensure that it's correct, and not all referees carefully read proofs. So, you should always be careful. Also, it takes so long for the refereeing and printing process to happen that quite often, what starts out as a reference to the arXiv will become a reference to a published paper, by the time your paper makes it to galley stage.

I would agree, though, that if a paper has been on the arxiv for a long time, but seemingly hasn't appeared in print, then you might want to be extra careful (and under such circumstances, I think actually writing something to warn the reader would be okay).