What I get from this is that during the process it passes through some version of gpt that is not aligned, or censored, or well behaved. So this internal process should not be exposes to users.