My interpretation is it quickly generates answers to keep it conversational but another process parses those messages for "prohibited" terms. Whether that second process is automated or human-powered is TBD