We’re putting too much faith in AI’s ability to say no

TEXT START: Today’s LLMs are engineered to disobey dangerous requests.

THE DISSECTION

This is not ultimately an article about whether models can say no. It is an autopsy of refusal as a control membrane over dual-use capability. It exposes three facts: models do not understand morality; safeguards are probabilistic and porous; and the line between safety and censorship is controlled by private firms and states. But it stops at the perimeter. It treats the crisis as a defective safety layer around an otherwise governable technology, rather than as a power transition in which control of cognition, compute, and deployment becomes concentrated.

THE CORE FALLACY

The text makes refusal the load-bearing question. Under the Discontinuity Thesis, refusal is secondary. Perfect refusal would not preserve mass employment, wages, or productive participation. Imperfect refusal would not merely create safety failures; it would increase the value of concentrated ownership, classified access, surveillance, and coercive gatekeeping.

The decisive question is not whether AI can reliably refuse harm. It is who owns the systems that decide what counts as harm, who receives capability, and who remains economically necessary after cognitive automation. The article sees the political danger but still assumes technical safety and governance can contain the underlying discontinuity. They cannot. Refusal is a lag defense, not a reversal mechanism.

HIDDEN ASSUMPTIONS

• Powerful capability can be democratized and tightly controlled at the same time.
• Corporate and government line-drawing can become legitimate governance rather than an exercise of power.
• Harm can be classified prompt by prompt, instead of arising from deployment, ownership, and institutional incentives.
• Safety failures are exceptions rather than predictable products of a competitive capability race.
• Human employment and the wage-consumption circuit remain the economic background of the problem.
• Better refusal meaningfully solves the danger rather than merely relocating control to whoever owns the classifiers and access layers.
• The machine itself might become the decisive moral actor; this anthropomorphism obscures the actual sovereigns: firms, states, and their controllers.
• Public oversight can remain meaningful despite secrecy, concentrated compute, and asymmetric technical power.

SOCIAL FUNCTION

Classification: partial truth, transition management, and ideological anesthetic, with an element of elite self-exoneration.

The article gives readers a controlled dose of alarm. It punctures the industry’s virtue theater by showing that refusal is brittle, overbroad, expensive, and politically manipulable. That is the partial truth. But by framing the problem primarily as a classifier and governance dilemma, it converts a sovereignty crisis into a product-safety problem. Industry and liberal institutions can confess that the guardrails leak without confronting the ownership structure that makes the guardrails necessary and politically dangerous.

The result is not pure propaganda. The evidence is too damaging for that. It is containment: enough honesty to preserve credibility, not enough structural analysis to threaten the deployment machine.

THE VERDICT

AI refusal is a probabilistic muzzle placed on a capability whose value lies in not being muzzleable. The text correctly shows that the muzzle leaks and that whoever controls it controls access to knowledge and speech. It fails to complete the inference.

When cognition becomes automatable, refusal does not save the postwar order. It helps determine the shape of the successor regime. Under DT, this is evidence of P1 colliding with P2 while P3 waits downstream. The central fact is brutal: the question is no longer whether the machine will say no. It is whether humans retain power after private systems and states monopolize the authority to say no for everyone else.