Fuel cells are electrochemical devices with some internal variables that cannot be measured, but have to be monitored in real-time. That is the case for the liquid water inside the fuel cell catalyst layer. This motivates the development of online algorithms, i.e., observers, able to estimate such variables. Nonetheless, fuel cell dynamics are strongly nonlinear, with significant parametric uncertainty and significant sensor noise. Therefore, typical observers, the extended Kalman filter (EKF), usually underperform or are unstable in such systems. To overcome such limitations, this work proposes a novel nonlinear observer to estimate the liquid water saturation in fuel cells based on an inherent differential detectability of the liquid water dynamics. The stability of the proposal is formally analyzed and it is validated through numerical simulations and in an experimental prototype, where noise and uncertainty are considered.