The fact that a theory is not a model to begin with. The fact, also, that theories are based on metaphysics principles that cannot be proven right but neither false. The fact that one can also complexify models of a given theory to obtain good predictions yet shitty explanation. The fact that it is also non trivial to link experimentations with models (both upward and downward modelization problems), etc etc.
See this answer Newtons law saved by ad-hoc hypothesis? for more details.
edit: Let me give more details about the upward and downward modelization problems. Given a theory T, a model M of such theory, the downward modelization problem stipulates that it is by no mean trivial to say that an experimentation "in real life" is indeed an instantiation of M. That is, how to "produce" (with a good amount of certainty) the model M by an experiment E, and claim that what one is indeed measuring is M within T? Couldn't E' be a more faithful realization of M, or is it even possible to realize M with some experiment?
The upward modelization problem is the reciprocal. Given an experiment E (with some data, hypothesis and so on), how to modelize it with a model M within T? How to be sure that M is indeed the "best model" describing E within T, and that M' wouldn't be better or more relevant? That is, what does "better model" means when one is trying to fit experimental data? Certainly, elegance and simplicity is more important than no error with high complexity and bunches of free parameters.