This is a separate mechanism. The user is not notified about the flagging and rather than redirecting to a weaker model, the response is intentionally sabotaged.

It's user-hostile to the point of parody.