This contribution discusses a method for approximating the observability canonical form of nonlinear systems, circumventing the need for extensive symbolic computations. Instead, we design a high‐gain observer leveraging neural networks and automatic differentiation. The approach aims to address the challenges associated with computing the observability canonical form, and especially the reverse transformation, by utilizing neural networks to approximate the nonlinearities in the observer's differential equation and the inverse observability map. We demonstrate the effectiveness of the method through experimental results on a physical pendulum system.