> and rewording it

Using the probabilities encoded in the training data.

> In that sense they are not compressing the data

You're right. In this case they're decompressing it.