It depends on the design. Maybe an AI would require learning cycles where it spends time processing it's experiences and making alterations to it's equivalent of neural connections and synaptic weights, like a neural net in training mode. For neural networks used in real-world applications, often significant time is spent on training the neural network and then the network is frozen and learning is disabled for speed, efficiency, predictability and reliability in the final product.
Maybe androids would dream of electric sheep.