A new proposal on the AI Alignment Forum argues that most alignment failures stem from poor value generalisation. The author claims this lack of generalisation is a fundamental reason why the alignment problem remains difficult. This framework redefines failure modes as generalisation gaps.