The AI safety rule workers need: no agent gets power without a named human owner

TEXT START: Britain is moving quickly to put artificial intelligence into real institutions.

The Dissection

The article is not really about AI safety. It is about making autonomous systems legible to existing workplace accountability structures. Its rule—every consequential agent must have a named human owner with power to stop, reverse, and review it—is a sensible control mechanism for limiting damage. But it treats governance discipline as if it were a solution to the deeper economic rupture.

The text converts a structural question—who controls productive AI capital?—into an administrative question—who signs for the machine? It offers workers a handle on liability while leaving ownership, deployment incentives, and the substitution of human labor untouched.

The Core Fallacy

The central error is confusing assigned responsibility with actual control.

A named human owner does not necessarily own the system, understand its internal reasoning, control its permissions, possess the authority to halt deployment, or have the resources to reverse its consequences. In many organisations, that person will be the nearest worker available when the system fails: accountable in public, powerless in practice.

Authority budgets, escalation paths, logging, and independent review are real safeguards. They are lag defenses against misuse and runaway action. They do not defeat P1, P2, or P3. Once AI agents are cheaper and more capable than human labour, institutions will face relentless pressure to expand their authority. Human approval becomes a bottleneck to be minimised, automated, or ritualised. The owner becomes a ceremonial brake attached to a vehicle whose economic purpose is to eliminate the driver.

Hidden Assumptions

• That a human can monitor and understand agent behaviour quickly enough to exercise meaningful control.
• That employers will grant the named owner genuine stop, reversal, and review authority rather than nominal responsibility.
• That legal and institutional accountability will remain aligned with operational control.
• That human approval requirements can survive competitive pressure once autonomous agents outperform human workers.
• That better safeguards produce broad worker protection rather than simply making automation easier to deploy.
• That the main labour danger is uncontrolled agent action, rather than the destruction of the wage-to-consumption circuit itself.
• That workers will retain meaningful bargaining power after their roles become economically unnecessary.

The article also assumes that “safe adoption” is a shared objective with shared gains. Under the Discontinuity Thesis, adoption safety can serve the Sovereign by reducing friction, liability, and reputational risk while accelerating the replacement of workers.

Social Function

Primary classification: transition management and partial truth.

Secondary function: ideological anesthetic and elite self-exoneration.

The article acknowledges a genuine operational danger: autonomous agents can cross authority boundaries, and least privilege, human escalation, monitoring, and incident review can reduce that danger. Its worker-accountability critique is valid within the old system.

Its anesthetic function is the frame. By presenting named ownership as “the AI safety rule workers need,” it implies that workers can remain protected through better procedures while the productive basis of their employment is being removed. It makes the transition appear governable through workplace policy, allowing deployers to say they are responsible without surrendering control or slowing the underlying substitution race.

The proposed rule may therefore protect workers from being scapegoated in individual incidents, but it does not protect their economic position. It can even accelerate deployment by granting institutions a compliance wrapper around increasingly autonomous labour.

The Verdict

The article identifies a real fuse and proposes a useful firebreak. It mistakes the firebreak for a cure.

Human ownership of delegated authority can contain accidents, assign liability, and delay reckless deployment. It cannot restore productive participation once AI severs employment from income. The named human owner is a temporary governance artifact: valuable as a brake, inadequate as a survival strategy, and vulnerable to becoming a scapegoat or rubber stamp. The article manages the corpse of the old labour order; it does not prevent its death.