Into something probable. The models that underlie these chatbots are usually overfitted, so while they usually don't repeat their training data verbatim, they can.