Does AI invent things and train itself in the process?

Training and use are two separate phases. During training, a model analyzes huge amounts of text and in the process adjusts billions of internal numbers, called parameters. Afterward, these numbers are frozen. In conversation, the model uses what it has learned, but does not change in the process. If a chat seems to remember earlier conversations, that is a memory feature that stores notes, not new training.

AI can still invent things. It recombines known patterns in new ways. That often produces something useful and sometimes something surprising. The same ability also leads it to invent facts that sound plausible but are false.

Companies now also use AI-generated data to train new models. This happens in a targeted and controlled way. There is no model that secretly improves itself overnight.