The model is an encoder-decoder, which encodes some text into a latent embedding, and can then decode it back into text. It’s a feature of the model itself.