Disentangling Corrigibility: 2015-2021 — AI Alignment Forum