The state-of-the-art feed-forward control of active hand prostheses is rather poor. Even dexterous, multi-fingered commercial prostheses are controlled via surface electromyography (EMG) in a way that enforces a few fixed grasping postures, or a very basic estimate of force. Control is not natural, meaning that the amputee must learn to associate, e.g., wrist flexion and hand closing. Nevertheless, recent literature indicates that much more information can be gathered from plain, old surface EMG. To check this issue, we have performed an experiment in which three amputees train a Support Vector Machine (SVM) using five commercially available EMG electrodes while asked to perform various grasping postures and forces with their phantom limbs. In agreement with recent neurological studies on cortical plasticity, we show that amputees operated decades ago can still produce distinct and stable signals for each posture and force. The SVM classifies the posture up to a precision of 95% and approximates the force with an error of as little as 7% of the signal range, sample-by-sample at 25Hz. These values are in line with results previously obtained by healthy subjects while feed-forward controlling a dexterous mechanical hand. We then conclude that our subjects could finely feed-forward control a dexterous prosthesis in both force and position, using standard EMG in a natural way, that is, using the phantom limb.