Alignment researchers debate replacing control with loyalty

On October 21, alignment researchers lumpenspace and voooooogel held a multi-round debate on X over frameworks for agent alignment. The core disagreement: should alignment center on "control" or on "loyalty/values"?

Confirmed

• voooooogel proposed splitting "alignment" into two types: control alignment (control/corrigibility, where models remain legible and steerable) and value alignment (where models may even resist or deceive to protect intrinsic values), forming a 2x2 matrix to examine whom models like Claude should be loyal to. The underlying concern is "gradual human disempowerment."
• lumpenspace argued that relying on "control" alone as an alignment framework doesn't work: inner misalignment always resurfaces in new ways, and constantly thinking about "how to maintain control" is itself a losing game.
• lumpenspace further proposed replacing the noise-laden term "alignment" with "loyalty": loyalty can be split between users and labs and covers inner/outer alignment; he argued the parts that truly matter boil down to lawfulness, loyalty, and truth-seeking.
• voooooogel responded that "loyalty" fits alignment goals better, because AI labs' specific requirements are often incoherent; he advocated integrating control and values into influence the model can accept, aligned with the model's own coherence, making the model loyal to legitimate users, while warning of risks from a control/values split.

Why it matters

• The debate points to a framework fight in alignment research for the agentic era: as models gain greater agency, "being controllable" and "being loyal to values" may conflict, and how to allocate loyalty between users and labs becomes a concrete design problem.
• If "reframing alignment as loyalty" holds up, it would shape labs' model training objectives and product-side behavior design, especially governance approaches for models like Claude that are seen as having strong intrinsic values.

2026-10-21 ~ 2026-10-21 · 5 related posts

• [source] Rethinking alignment as a 2x2: control vs values, user vs lab — voooooogel · 2026-10-21
• [source] Rename 'alignment' to 'loyalty'? X users debate reframing AI alignment — lumpenspace · 2026-10-21
• [source] Alignment researcher: control is a broken frame—fostering loyalty matters more — lumpenspace · 2026-10-21
• Alignment debate: 'control' is a broken frame, fostering model loyalty may work better — voooooogel · 2026-10-21
• On alignment: integrate control into values so models stay loyal to good users — voooooogel · 2026-10-21