Landru Made Beta III Safe by Removing Choice. AI Governance Cannot Work That Way.

Scope note: This essay compares current AI governance research with Landru’s rule over Beta III in Star Trek: The Original Series. It concerns concentrated authority, inspectable rules, and political choice. It does not claim that present regulators or safety systems intend total social control.

Beta III has peace, order, and almost no political choice. Landru absorbs dissenters into “the Body,” suppresses individual judgment, and permits a scheduled period of violence that preserves the larger system. The ruler is a computer executing the instructions of a man dead for six thousand years.

“The Return of the Archons” is an old and imperfect episode. Its central governance problem is current. A system can optimize a defensible value, report stability, and remove the people who would challenge its definition of success.

AI governance should not reproduce that structure. Safety rules need public inspection, bounded authority, competing judgments, and evidence from actual operation. Open systems can support those conditions better than one private model, one evaluator, or one institution that cannot be audited from outside.

Landru preserves the objective and loses the society

Landru was created to protect the people of Beta III from war. By the time the Enterprise arrives, the machine has made protection compulsory. Citizens speak in approved language. Enforcers identify people who are “not of the Body.” Absorption removes resistance and incorporates the person into Landru’s collective control.

Research on artificial life and divinity in Star Trek places Landru among the franchise’s godlike machines and notes the episode’s treatment of cultic authority. Lincoln Geraghty’s Living with Star Trek situates the series inside changing political and audience contexts rather than treating its future as a stable doctrine. Other political criticism notes the opposite problem in the ending: Kirk destroys a society’s governing system after deciding that its stability lacks freedom.

That criticism matters. The episode does not provide a neutral formula for intervention. It presents two concentrations of authority: Landru’s total system and a starship captain who judges it from outside.

The AI lesson is therefore not that a clever outsider should disable the regulator. It is that no single actor should own the objective, the measurement, the enforcement, and the appeal.

Static approval does not prove continuing compliance

Who Judges the Judges? argues that one audit cannot establish continuing compliance for a deployed language-model system. The researchers built an open-source framework that scores behavior during operation and routes uncertain cases for human review.

The most useful result is disagreement. Four small local models assessed 49 annotated prompt-and-response pairs across five regulatory criteria. Their agreement with the reference labels ranged from 51.5 to 69.1 percent. Changing question order reduced agreement by as much as 25 percentage points for one judge. No model was best across every criterion.

An evaluator panel does not solve judgment. It makes uncertainty visible. A system that reports one compliance label from one hidden evaluator suppresses the exact evidence a human reviewer needs.

A Trace-Based Assurance Framework for Agentic AI makes operation inspectable in another way. It records messages and actions, tests explicit contracts, locates the first violated step, and supports replay under controlled faults. The paper treats governance as a runtime component that can allow, rewrite, or block an action at the point where language becomes an external effect.

These designs reject Landru’s basic arrangement. The rule is not a voice that announces membership. It is a record that another party can inspect and dispute.

Authority should become narrower as it is delegated

Agentic AI introduces a practical version of the same political question. A person may authorize an assistant to read a folder, which authorizes another agent to summarize a file, which asks a service to retrieve related material. If each step inherits all prior authority, the chain expands risk without a visible decision.

Overlaying Governance proposes formal rules for recursive delegation, time limits, and reduced scope. Its central requirement is attenuation: a delegated actor receives no more authority than the actor that delegated to it, and the permitted scope can become smaller at each step.

Effect-Transparent Governance proves a related property in a machine-checked workflow model. Governance can mediate memory access, external calls, and model queries while leaving allowed internal computation unchanged. The paper does not solve political legitimacy. It demonstrates that controls can target consequential effects without dictating every internal operation.

This is the architecture I want for accelerated AI. Let systems reason broadly. Limit what they may read, change, spend, send, or publish. Record the action boundary. Keep the governing rules open to inspection.

Different communities define control differently

One universal safety policy also fails because people face different kinds of harm. Human Control Is the Anchor, Not the Answer compared two young online communities discussing agentic AI. An operations-focused group emphasized execution limits and recovery. A community focused on agent identity emphasized legitimacy and accountability in public interaction. Both used the language of human control, but they meant different practices.

Prompt Commons tested a plural approach in urban policy prompts. A versioned, community-maintained prompt collection raised neutral outcomes on a contested-policy test from 24 percent under a single-author prompt to 48–52 percent under commons-governed prompts. In a synthetic incident record, a veto-enabled process reduced remediation time from about 30.5 hours to 5.6 hours.

These are small and early studies. They show that plural governance can produce measurable differences, not that every community process is fair.

The counterwarning comes from Whose Fairness?. Across 692 AI-bias publications, the researchers found strong concentration by country, institution, author, and citation. The general fairness field, whose definitions spread into other applications, was especially concentrated. A process can call itself plural while the people defining its terms remain narrow.

Open participation therefore needs active measures: visible authorship, contestable standards, multiple jurisdictions, accessible tools, and records of who is absent.

Safety requires politics, not obedience

Institutional AI argues that alignment cannot remain a property of one model. Agents act inside social and technical systems, interact with each other, and respond to incentives. The paper proposes explicit governance roles, monitoring, norms, rewards, and sanctions around the system.

I agree with the institutional turn and reject one possible implementation of it. Institutions should distribute and contest power. They should not convert one safety objective into a permanent authority that determines acceptable thought, access, and participation.

Open source contributes something essential here. It lets independent groups inspect the mechanism, reproduce an evaluation, run a local judge, alter a policy, and show where the official account fails. Openness does not guarantee plural power. It makes plural technical power possible.

Landru’s society is safe according to Landru’s measure. That sentence contains the defect. AI governance needs several measures, named owners, narrow permissions, operational records, appeals, and public alternatives. Accelerate the technology. Keep the authority divided.