Keras softmax layer returns one-hot encoded array

Question

I am using a simple net for classification but the output of the predict is one-hot encoded. I checked inside the loss function and the actual output is correct.