Public briefing · Not the complete paper
The machine may win the argument.
It does not thereby acquire the mandate.
The strongest case for machine benevolence begins with a real danger: a moral theory can become so impartial that it stops noticing the people standing directly in front of it. Dan Hendrycks names that danger clearly. This paper accepts the warning—and then asks what happens when the proposed cure is itself allowed to cross a boundary without permission.
The answer turns on a distinction that intelligence cannot dissolve. Better judgment is not the same thing as legitimate jurisdiction. A system may identify a superior outcome, explain it flawlessly, and predict that humans will later approve. None of those achievements manufactures the authority to act now.
Human agency would be a conditional privilege if retaining it required outarguing a superior intelligence.
Does the safeguard still bind?
A protection is not preserved merely because its name survives. Change the mathematics at the boundary and the old guarantee may disappear while the rhetoric remains intact.
Who authorized the effect?
Evidence may inform a decision. Inference may enlarge the option set. Neither may quietly create the permission required to cross into protected consequence.
Who authorized the rule-maker?
Audits can expose compromised governance. They do not automatically grant the auditor coercive power. Scrutiny itself receives no exemption from scrutiny.
The finding · shown, not advertised
The wall became a price.
Eigenism’s logarithmic branch associates divergence near zero with protection of an irreplaceable carrier. Its optional extension to negative wellbeing shifts the logarithm. At the very boundary invoked by the protection, the two expressions do not say the same thing.
The singularity can make zero non-compensable under the paper’s stated finite conditions.
The singularity is gone. The same boundary now has an ordinary finite score.
Proposition 1 · two carriers
For positive weights α and β, let
S(w1,w2) = α ln(1+w1) + β ln(1+w2).
Compare A = (x,y), with x > 0 and y ≥ 0, against B = (0,z). Then the shifted objective prefers the alternative that drives the first carrier to zero exactly when
z > (1+y)(1+x)α/β − 1.
What this establishes. Once the singularity is removed, the protected zero is no longer a wall inside this objective. A sufficiently large gain elsewhere can outbid it.
What it does not establish. It does not equate every wellbeing zero with death, lost consent, or a rights violation; it does not refute every Eigenist objective; and it does not diagnose anyone’s intent.
The narrow charge is harder to evade: a representation cannot keep receiving credit for a guarantee that its revised mathematics no longer supplies.
A scheduling assistant may read availability, infer preferences, and draft the perfect invitation. If its grant excludes send, accuracy does not add the missing verb. The same principle scales upward. Capability is not jurisdiction. Confidence is not consent. Benevolence is not authority.
The full paper then turns upstream: six Moral Covering of Desire gates test how a mandate is formed and amended; a documentary pilot examines Anthropic’s Responsible Scaling Policy v2.0 without pretending that a documentary score is a license to intervene; and bounded executable evidence tests effect-time authorization, revocation, replay, retargeting, and negative controls.
The test of a boundary is whether it still binds when the actor it constrains believes that crossing it would do good.