We propose a novel method for movement assistance that is based on adaptive oscillators, i.e., mathematical tools that are capable of extracting the high-level features (amplitude, frequency, and offset) of a periodic signal. Such an oscillator acts like a filter on these features, but keeps its output in phase with respect to the input signal. Using a simple inverse model, we predicted the torque produced by human participants during rhythmic flexion-extension of the elbow. Feeding back a fraction of this estimated torque to the participant through an elbow exoskeleton, we were able to prove the assistance efficiency through a marked decrease of the biceps and triceps electromyography. Importantly, since the oscillator adapted to the movement imposed by the user, the method flexibly allowed us to change the movement pattern and was still efficient during the nonstationary epochs. This method holds promise for the development of new robot-assisted rehabilitation protocols because it does not require prespecifying a reference trajectory and does not require complex signal sensing or single-user calibration: the only signal that is measured is the position of the augmented joint. In this paper, we further demonstrate that this assistance was very intuitive for the participants who adapted almost instantaneously.