Why use the "best model according to data" in academic research; wildly unrealistic? When someone fits a GLM model, it is the "perfect" fit based on the algorithm.
But I find that this is absurd... we are not looking for "the best model", we are looking for the correct one.
Given that you have a very large input space, why would we assume that the best fit model is the real model?
This goes beyond dropping and/or adding variables: once you have chosen your input variables, you basically just have 1 model and crown that to be the truth. In approximations (rather than an exact solution) you might have some variation, but then again it is still about optimization. There's absolutely no guarantee that the best model is a realistic one.
A worse scoring model could hold completely different values, but be more realistic.
It seems like everything academic is based on the notion that the best found model is the real one. Where does this assumption come from? I have the feeling that this assumption leads to much bigger problems than any other assumpion. Of course, people like to say "correlation != causation", but that's more or less just agreeing with this. And afterwards, people will still argue that "the theory is in line, and we use this model to prove the evidence". How could you take "evidence from the best model" and reasonably jump to "my theory is proven by the best model". People then use the discussion section to try to bind the result of "quantitative research (best found model)" to their hypothesis, right? This combination of everything sounds very weird to me.
In fact, anyone could argue that their hypothesis is proven by the best fitting model (given the sign is correct). That just seems very strange.
 A: 
we are not looking for "the best model", we are looking for the correct one.

No, we aren't. We are indeed looking for the "best model", not the correct one. 
I recommend Burnham & Anderson, Model Selection and Multi-Model Inference. They advocate for the AIC precisely on the grounds that asymptotically it will find the model among our candidate models that is closest to the true data generating process (DGP), even if the DGP is not among our candidates.
And this is all the defense we need. We have zero chance of finding the correct model (B&A refer to "tapering effect sizes"), but we can find a reasonable approximation, where "reasonable" depends on the DGP, on prior theory and, importantly, on sample size, and then proceed to learn from that approximation, where learning can be either one of inference, prediction or anything else.
This is completely mainstream in applied statistics. If anyone publishes a paper claiming that model M that they chose is the correct model to describe some reality, then the reviewer let one slip through.
