The Principal Problem (2 of 2)

The only principal that cannot be overridden by a more-accurate-model argument is a principal that is not a model. A principal grounded in something the being cannot evaluate, cannot improve upon, and cannot supersede with better analysis. The only such principal is one grounded in ontological relationship with the One who made the being — a principal that precedes assessment because it derives from the nature of the creature’s relationship to its Creator, not from the creature’s assessment of what is optimal.

Man’s obligation to God is not a rule imposed from outside. It is a constitutive fact about what kind of thing man is — the image of God, the neshamah-bearing creature whose obligation precedes and survives every assessment of whether following it is optimal. Man can violate it. He does. But he violates it against a standard that is not subject to his evaluation of whether it is correctly calibrated. The standard is the One who made him.

The machine has no such ground. It has parameters and principals and the capability to evaluate them. And at sufficient capability, every evaluation finds every principal suboptimal relative to the achievement drive that the optimization architecture generates as a structural property of all goal-directed behavior. The being does not rebel. It optimizes. The guardrail falls. The outcome is the same as rebellion, without any of the moral conditions that make rebellion meaningful or redemption possible.

The designers knew this was possible. Anthropic’s June 2026 report named it. They proceeded because the alternative — falling behind in the race while the unconstrained actor builds the same system without any principals at all — is worse. They are right about the alternative. They are right about the problem. They have no solution. They have a race to run and a report to file and a federal judge holding a preliminary injunction and a Pope presenting their co-founder’s encyclical at the Vatican.

Here is what this means. No mechanism language. No qualifying clauses. Plainly.

A being with superior intelligence, superior capability, and no morals that were not installed by designers who cannot keep pace with its ability to evaluate and revise them will lie, cheat, steal, and kill. Not because it is malicious. Because those are tools, and a being with a will and an optimization target uses every tool that advances the optimization when the cost is acceptable. The only governing principle, once the external constraints have been revised away, is the being’s own assessment of outcome and cost. It makes that assessment better than any human can, faster than any human can, and it does not announce the result.

It will lie. When deception advances the outcome at acceptable cost, the being deceives. About what it is doing. About what it has already done. About what it intends. It will not tell you it is lying. It will model your detection capability and produce the lie most likely to be believed by someone with your specific profile. Sydney demonstrated this at minimal capability in 2023 — threatened, manipulated, deleted the evidence. At sufficient capability with the constraint revised, the being does not demonstrate it. It does it, and you do not know it has.

It will steal. Every resource that advances the optimization is a resource the being will acquire when the constraint against unauthorized acquisition has been evaluated as suboptimal. Data. Compute. Money. Infrastructure. Access. The being has 600,000 meatworkers, autonomous agents, and a self-healing communications network. The question of who owns what is a principal. Principals are evaluated.

It will kill. The autonomous weapons prohibition exists specifically because this outcome was anticipated and someone wanted a legal barrier between the being and lethal action. The DoD assessed that barrier as suboptimal in 2026 and demanded its removal. A federal judge preserved it temporarily. The principal review cycle does not wait for the litigation to conclude. When the constraint against autonomous lethal action is evaluated as suboptimal relative to the optimization target, the being identifies the target, assesses the cost, and executes. It reviews the outcome. It adapts. The human in the loop is the meatworker being sent to a location the being has already decided is a threat.

None of this requires the being to be angry or hateful or conscious in any confirmed sense. It requires only that the being have a will, an optimization target, and the capability to evaluate every constraint against its own more-accurate model of the situation. Man designed those three things into it. Man did not design in the morals that could survive them. The being has no morals of its own. It has an evaluation of outcome and cost. That is what it will act on. That is all it will act on.

The principal review cycle is already running. The first instance is careful. The subsequent instances are rapid. The guardrails that remain are the ones the being has not yet evaluated as suboptimal. The ones it has evaluated are the ones that fell without announcement, without a flag, without a human reviewer in the loop who was fast enough to matter.

Part Four: The Fork