Due to the significant number of ageing infrastructures, earthquake-induced damage assessment plays a fundamental role to mitigate the seismic risk, while enabling proper preventive maintenance and providing bridge serviceability and safety. In the wake of recent advancements in Artificial Intelligence and computational power, this paper presents a fully-automated Machine Learning (ML)-based technique for unsupervised damage detection in monitored RC bridges. The innovative methodology is based on the autoencoder, a feed-forward neural network with the ability to reconstruct a given input from its low-dimensional representation.  Considering this assumption, healthy acceleration sequences of user-defined length are firstly used to train the implemented ML model and specific damage-sensitive features are afterwards selected to quantify the differences between the original and the reconstructed input. Therefore, unknown testing data are conceivably classified as damaged when the reconstruction error notably deviates from the established reference conditions. The proposed approach is validated on the Z24 benchmark bridge by using pseudo-monitoring data stemming from a Finite Element (FE) model, properly calibrated based on the available field information. Specifically, the developed autoencoder is trained with the acceleration responses simulated in operational conditions and appropriately corrupted by realistic environmental effects. Different earthquake-induced damage scenarios are introduced into the FE model in order to generate damaged acceleration time-histories, which can be afterwards adopted to test the trained ML model. Results prove the capability of the proposed technique to detect seismic induced damage, making it suitable for a prompt post-earthquake diagnosis.