Researchers have usually advanced mechanistic plasticity modelling with novel formulations that lack rigorous analysis of model error sources. Recent findings have demonstrated that models with non-unique parameters cannot capture univocally the local response at or below the grain scale. Moreover, optimisation approaches, including artificial intelligence, that fit parameters to the macroscopic responses (i.e. stress-strain curves) are ineffective for shallow objective functions with multiple minima. These approaches do not address the lack of testing data that hinders parameter identification; the alternative lies in enriched dataset fuse data from across scales and sources. This talk will discuss sources of uncertainty in crystal plasticity modelling under monotonic and cyclic loading. We will present different examples of epistemic uncertainty such as model form and model input errors, numerical errors or biased calibration and validation. We will also cover examples related to aleatory uncertainty not related to materials' intrinsic variability. We will conclude by examining opening challenges and opportunities to advance the prediction of mesoscale responses in metals.