My understanding from psych class is that it doesn't work in the way people usually think of.
However, there is something called "priming." If you see a word like "duck" and then are asked to read aloud a second word, you'll read "quack" faster than "bark". This is true even if the word "duck" is presented so quickly that you aren't aware of having seen it. You could even present the word "duck" aurally.
Some labs have lately been claiming evidence of more complex priming. For instance, if you spend some time thinking about your mother, you'll do better on a trivia came, supposedly because thinking about your mother makes you want to do well. A lot of this work has been done at Yale:
http://www.yale.edu/psychology/FacInfo/Bargh.html.
There's a lot of evidence for it, so I supose it's true, but I still have trouble believing it. One of these days, I'll probably try to replicate some of those findings at my Web-based lab:
http://coglanglab.org.
But as far as fitting in a message like "smoke less" into the background of a Mariah Carey song, I don't think there's evidence that it works.
----
http://coglanglab.org