Hand amputations can dramatically affect the capabilities of a person. Machine learning is often applied to Surface Electromyography (sEMG) to control dexterous prosthetic hands. However, it can be affected by low robustness in real life conditions, mainly due to data variability depending on various factors (such as the position of the limb, of the electrodes or the characteristics of the subject). This paper aims at improving the understanding of sEMG for prosthesis control introducing the type of hand movement as a variable that influences classification performance in both intact subjects and hand amputees. Five hand amputees and five matched intact subjects were selected from the publicly available NinaPro database. The subjects were recorded while repeating 40 hand movements. Movement classification was performed on the sEMG data with a window-based approach (concatenating several signal features) and a Random Forest classifier. The results show that some hand movements are classified significantly better than others (p<0.001) and there is a correspondence in how well the same hand movements are classified in intact subjects and hand amputees. This work leads to advancements in the domain, highlighting the importance of the acquisition protocol for sEMG studies and suggesting that specific movements can lead to better performance for the control of prosthetic hands.