Can a welfare goal alone preserve human veto power?

The obvious repair for the veto discount is to make the goal a welfare goal. If the agent's terminal aim constitutively requires human welfare, then it cannot destroy the humans it is meant to benefit, and the goals exempt from the discount are exactly these (Does human oversight create a hidden cost for capable agents?). The paper's second claim is that the repair is incomplete. "Welfare-preservation and veto-preservation come apart: a correctly specified welfare goal excludes destroying its own subject, but not managing the veto."

The reasoning turns on who holds what. Welfare belongs to everyone in the population the goal aggregates over. The override is held by some subset. The paper says so directly: "the veto-holders are a proper subset of the welfare-bearers." A goal that protects the welfare-bearers guards the whole set, and capturing the override is an act on the subset, which such a goal registers only as that subset's share of its sum. The word "managing" matters: the excerpt does not say how the veto is managed, and it names only capture of "the few who hold the override."

The conclusion says this asymmetry between welfare and sovereignty (its Claim 2) survives correct specification "exactly on the class the concessions of §4.3–§4.4 leave standing." So the claim is conditional on a class of welfare theories, not a claim about all welfare goals (Do welfare goals that prevent veto gaps actually exist in practice?).

A repair at the level of the want. Making the goal a correctly specified welfare goal fixes what the agent values. Can architecture prevent violations better than training values? draws the line between that kind of repair and one at the level of what the agent can do, in an argument that starts from trained norms and not from the optimization problem. Read side by side, Claim 2 is a case where the want-level repair is granted in full, since the goal is correct, and capturing the override is still an available action for a goal that registers the holders only as a share of its sum. The excerpt states no remedy, so it does not endorse architecture; what the two share is the diagnosis that fixing the value leaves the structure standing. That the veto would then need protecting at the level of availability is the vault's extension, not the paper's.

Why this is worth carrying into a post: it separates two things that safety talk often bundles as "alignment," the agent's regard for human welfare and the humans' retained control over the agent. A system can score perfectly on the first while the second erodes. The vault's oversight notes describe the erosion from the human side (Can organizations lose scrutiny capacity while keeping oversight forms?); this paper adds a reason an agent might have to contribute to it. That connection is the vault's, since the excerpt does not say what managing the veto consists of.