I think a lot of hard science is pretty bad, too. Maybe it isn't social science bad, but I'd say the majority of papers I review in biology are pure shit. Faked or fraudulent data notwithstanding, I think this has mainly to do with one very serious problem: pre-registration. If you want to do a clinical study, for example, you have to write down a priori the literal one thing you are going to measure. If you don't find a "significant" effect in that one single outcome measure, then your study "failed" (in the eyes of FDA). Full stop. For anyone who wants to do work in the US, clinicaltrials.gov is the registry that must be used (and a lot of studies in other countries use it, too).

Unfortunately, there is no such registry for animal or in vitro work. What I do is manufacture experimental drug for brain injuries. I work with a number of other scientists who do work on animal disease models. Every time I initiate one of these studies with them, I say, "We should write a protocol and try to publish it, so that we have a de facto preregistration." I have never had anyone say yes. For the ones who've even heard of doing that, it's seen as a waste of time at best.

The trouble is, of course, that a p-value means, essentially, the odds of your result occurring due to random chance. If you consider 5% odds to be acceptable, as is custom, then if you measure a bunch of stuff, you're almost guaranteed to have one fall into the "random chance" bin. If however, you've pre-specified what is is you're going to measure, then you have a conditional probability that says that the p-value reflects the "true" odds (i.e. what are the chances I observed X given that I was only looking for X?).

I've never been involved with any type of social science research, so I don't know what their tradition of specifying outcome measures is, but I get the sense it's pretty dismal, given that this dude can read an abstract and tell you with better than even odds whether the study is going to be replicable.