How do you know its a random users message and not a generated user message based on the training data?