This is great work:
"As expected, across these experiments, the researchers found that the more people smiled, the happier they reported being. But only some people. Surprisingly, for a section of the population, smiling actually reduced well-being. The more these people smiled, the less happy they were. This is like finding that there are some diners who, after consuming a four-course meal, feel less full!"
It makes perfect sense to me, and it's related to what I told The Chronicle of Higher Education a while back, that some people don't dispositionally approach happiness the same way. This finding is more at the implicit and affective level, but I wouldn't be surprised if it corresponded to a different broad disposition toward happiness. For example these might be high-meaning people, less affectively driven.
It also something I plan to add to my Media Tips page: Many main effects are misleading. Some main effects are false for the majority of the population, meaning that an effect of the form X causes Y will often only be true for a minority of people. That's not the case here, but we still see a significant population for whom a famous effect does not hold, and in fact is reversed. Significant correlations between X and Y do not require that X predict Y in the majority of cases, or that being high in X (assuming a continuous variable now, like a personality trait) means one is likely to be high in Y. A significant correlation does not mean that at all. It can be driven by 10% of the sample.
I've developed some new statistics, new coefficients, to help clarify what a correlation actually represents in many cases. These will be detailed in some upcoming journal articles, hopefully. It would apply nicely to the example above, and the pencil-in-the-mouth experiments.