LessWrong Paper Introduces "Self-Unalignment": AI Subagents Drifting from Their Own Values

A new LessWrong paper by Jan Kulveit and Rose Hadshar introduces the "self-unalignment problem," a failure mode where AI subagents quietly drift from their own values. Related discussion highlights the multi-principal dilemma between user and vendor interests in AI alignment.

2026-10-27 ~ 2026-10-27 · 2 related posts

• Aligning with Microsoft is ambiguous: Sydney's multiple-principal problem — jankulveit · 2026-10-27
• New LessWrong essay on "the self-unalignment problem" in AI alignment — jankulveit · 2026-10-27