They have to realize that the language model did it's job just fine. It produced a convincing piece of well-formed language that fits the context window. If they wanted knowledge instead, they should have used a knowledge model.
I hope you do realize that "artificial conman" is hardly a goal.