I guess the point is that if there are a large number of big, variable, competing effects with interactions etc, you can't just look at each of them in isolation and expect them to predict a large fraction of the variance. So you need a more sophisticated (possily impossibly complex) model or study, as you said.
In the nudging example, this appears to imply something like "conditioned on a large effect having been measured using a sufficient amount of data (!), you have to be extremely confident in the study design and model to actually believe it"?
