Reconstruction loss is a measure used in machine learning to quantify the difference between the input data and its reconstruction from a model, often used in autoencoders and generative models. It serves as an objective function to guide the training process, ensuring that the model learns to accurately capture the underlying patterns of the input data.