When Compression Becomes Authority

A recurring fantasy in contemporary AI politics is that bureaucracy can be made obedient.

The story goes like this: institutions are slow because humans get in the way. Civil servants distort commands, preserve their own power, introduce delay, hide behind process, or refuse to execute the will of leadership with sufficient enthusiasm. Bureaucracy becomes a bad routing layer between sovereign intent and administrative reality. AI then appears as a bypass: a machine that can receive instructions from the top, process the relevant facts, and carry out policy without all that human friction.

This is not a theory of administration. It is a vending-machine theory of power.

Insert intent. Receive outcome. Shake if necessary.

The problem is not merely that this fantasy overestimates AI. It also misunderstands bureaucracy. Bureaucracies are not just command pipes. They are systems for reducing complex social reality into administrable form: categories, records, offices, eligibility rules, queues, notices, appeals, signatures, budgets, exceptions, precedents, and files. They are often maddening. They are often unjust. They are often self-protective, indifferent, evasive, and cruel in that distinctively procedural way that makes people fantasize about setting a copier on fire.

But they are not just friction. They are also memory, category maintenance, error absorption, and legitimacy machinery. They store past failures in rules. They route disagreement through offices. They make certain kinds of consequence harder to impose without a record. They do not eliminate power. They format power.

AI does not remove that formatting. It changes where the formatting happens.

Henry Farrell and Cosma Shalizi have been arguing for a better frame: AI models should be understood not as baby gods, proto-agents, or Singularity countdown clocks, but as social technologies. Large models mediate access to the thoughts, words, classifications, and patterns produced by other humans. They do not stand outside society. They reorganize social material.

If AI is a social technology, then administration is one of the places where that claim becomes operational. It is where mediation stops being a metaphor and starts deciding who gets routed, delayed, denied, flagged, noticed, reviewed, remembered, or ignored.

That matters because AI enters institutions that already run on abstraction. Markets compress social complexity into prices. Democracies compress political conflict into votes, districts, parties, and mandates. Bureaucracies compress lives into cases, forms, statuses, and procedures. Large AI models add another compression layer to systems already built from compression.

The political question is not whether compression can be avoided. It cannot. The question is what happens when compression gains authority.

Bounded rationality, now with procurement

Herbert Simon remains useful here because he refused the fantasy of unlimited cognition. Human beings do not optimize against the full complexity of the world. They search under constraints. They satisfice. They build organizations because no single mind can model all relevant facts, alternatives, tradeoffs, histories, incentives, and consequences.

This is the part people often flatten into a slogan: bounded rationality. But the point is not just that individual humans are limited. The point is that institutions are architectures for living with those limits. Bureaucracy is one such architecture. It decomposes problems, assigns roles, stabilizes categories, and makes large-scale action possible by deciding what counts as relevant at each step.

That is both necessary and dangerous. The same category that makes a system administrable can erase the thing that matters. The same rule that prevents arbitrary discretion can produce mechanical cruelty. The same record that preserves accountability can become a substitute for reality.

AI belongs inside this problem, not above it.

AI does not abolish bounded rationality. It changes the architecture of boundedness. It allows institutions to summarize more quickly, classify more extensively, rank more cases, draft more explanations, surface more patterns, and act through more interfaces. Some of that will be useful. Some of it will be grotesque. Most of it will be both, which is how administration usually works when it has discovered a new toy.

The danger begins when institutions treat AI-generated compression as cleaner than older bureaucratic compression because its politics are harder to see.

A caseworker’s discretion can be challenged, at least in theory. A rule can be published. A form can be inspected. A file can sometimes be requested. A supervisor can be named. These systems are imperfect, often brutally so, but their seams are familiar.

AI-mediated discretion hides in different places: model outputs, vendor systems, training data, prompts, confidence scores, generated explanations, workflow defaults, dashboard categories, review queues, and interface design. The system appears procedural because it produces artifacts that look like procedure. It summarizes. It classifies. It recommends. It explains. It records.

This is how counterfeit legibility enters the institution.

The machine does not merely produce an output. It produces an output in a form that bureaucracy can metabolize.

The admissibility problem

The central question for AI-mediated administration is not whether the model is intelligent. It is not even whether the model is accurate in some broad benchmark sense.

The question is: what is this output allowed to become?

A model summary is not the file. A classification is not a decision. A recommendation is not an authorization. A generated explanation is not necessarily the reason anyone acted. A receipt is not proof that the action was legitimate. A persistent record is not fresh evidence merely because it can be retrieved later.

The institution has to decide what administrative role an AI output is allowed to perform.

This is the admissibility problem.

An AI output might be admissible as observation support: extracting information from documents, summarizing intake, transcribing audio, or flagging missing fields. That does not make it admissible as interpretation.

It might be admissible as interpretation: assigning a tentative category, routing a case, or identifying possible eligibility issues. That does not make it admissible as recommendation.

It might be admissible as recommendation: suggesting review, escalation, deferral, approval, or denial. That does not make it admissible as authorization.

It might support authorization. That does not mean it owns consequence.

It might record what happened. That does not mean it justifies what happened.

It might persist in institutional memory. That does not mean it should govern future decisions.

These distinctions are not academic fussiness. They are the difference between assistance and laundering.

Compression-authority laundering occurs when a lossy abstraction enters an institution as convenience and acquires the practical standing of fact, justification, or command without a separate admissibility process.

It rarely arrives wearing a little villain cape. It happens through workflow pressure.

A summary is faster to read than a file. A risk score is easier to route than ambiguity. A generated rationale satisfies the notice template. A dashboard gives managers something to measure. A vendor system promises consistency. A human reviewer clicks through because the queue is impossible and the recommendation looks plausible.

At no single moment does anyone necessarily declare that the model is now the decision-maker. The escalation is quieter than that. The output becomes the coordination surface. Once the organization depends on the surface, it begins to defend the surface.

That is when a proxy becomes jurisdiction.

Sense, authorize, act, persist

The easiest way to see the problem is to split AI-mediated administration into four stages: sense, authorize, act, and persist.

Each stage contains smaller rungs where laundering can occur.

Sense

Sensing is how the institution makes the world available to itself. It includes observation and interpretation.

Observation is what was submitted, measured, reported, recorded, transcribed, extracted, or summarized. Interpretation is the assignment of meaning: urgent, routine, eligible, noncompliant, suspicious, threatening, low priority, high risk, needs review.

AI systems are very tempting at this stage because the volume problem is real. Institutions have too many files, too many claims, too many messages, too many cases, and too little attention. A system that can summarize and classify at scale looks like relief.

Sometimes it is relief.

But sensing does not authorize consequence. A summary can omit. A transcription can distort. A classification can import category errors. A risk label can become sticky before anyone asks whether the system had the authority to assign it.

The rule should be simple: sensing may shape attention, but it does not settle standing.

Authorize

Authorization is where recommendation meets power.

A recommendation proposes what should happen next. It can be useful precisely because it is not yet a decision. The danger is when recommendation is treated as authorization without the standing transition that would license it.

This is the native failure mode of AI bureaucracy.

The system recommends denial. A human clicks accept. The institution later insists a human made the decision. Formally, perhaps that is true. Substantively, it may be nonsense. The human has become a ritual checkpoint, not an accountable decision-maker.

Authorization requires standing. The authorizing actor or process must have legitimate authority over the relevant operation, target, basis, and consequence. A model may classify a case accurately and still have no standing to authorize a benefit denial, account suspension, immigration consequence, fraud referral, personnel action, or police dispatch.

Authorization also requires consequence ownership. Someone must own the burden of being wrong.

If the agency blames the vendor, the vendor blames the configuration, the reviewer blames the queue, the manager blames the policy, and the policy points back to the system, authority has not been delegated. It has been orphaned.

Act

Action changes the world.

Benefits are denied. Accounts are suspended. Services are delayed. Workers are disciplined. Permits are rejected. Police are dispatched. People are flagged, excluded, investigated, charged, removed, ignored, or made to wait.

The admissibility burden rises with consequence. Low-consequence, reversible actions can tolerate more automation. High-consequence or hard-to-reverse actions require stronger evidence, clearer standing, and contestable authority.

Administrative convenience is not admissibility.

This sounds obvious until a system has a backlog, a budget target, a leadership priority, a vendor contract, and a dashboard showing improved throughput. Then convenience develops a theology.

Persist

Persistence is where AI-mediated administration becomes self-confirming.

Receipt records what happened. Contestation determines whether the affected person can challenge the decision chain. Memory determines what survives into future decisions.

This is the stage that looks like an afterthought and is probably the sleeper core of the whole problem.

AI-mediated administration does not merely remember its errors. It can metabolize them into future administrative reality.

A bad summary becomes a durable record. The durable record becomes a future feature. The future feature informs a later classification. The later classification appears to corroborate the original category. Eventually the institution encounters its own residue as evidence about the world.

The error is no longer an error. It is context.

This is memory poisoning.

The fix is not simply deletion, though deletion may be necessary. The memory layer needs provenance, decay, revalidation, and contestable custody. An output that was admissible as a draft summary in one workflow should not remain admissible as durable evidence in another. Persistence changes the role of an output. Role change requires renewed admissibility.

A receipt is evidence that something happened. It is not evidence that the thing was authorized. A stored classification is evidence that the institution once classified someone. It is not proof that the classification was valid, current, or usable for future consequence.

Without that distinction, the administrative record becomes a compost heap with subpoena formatting.

Counterfeit legibility

AI systems are especially dangerous in administration because they are fluent.

Older automated systems often produced scores, flags, or rigid outputs. Those were dangerous enough. Large language models can also produce explanations. They can generate individualized-sounding reasons, policy language, summaries of evidence, polite notices, appeal responses, and internal memos. They can make a decision look reasoned even when no accountable reasoning occurred.

That is counterfeit legibility.

The institution receives the benefits of reason-giving without the burden of reasoning. The notice is coherent. The file is populated. The workflow is complete. The dashboard is green. The decision has an explanation-shaped object attached to it.

But explanation-shaped is not the same as explained.

This matters because many administrative systems depend on rituals of reason-giving. A denial notice, disciplinary letter, fraud referral, moderation decision, or eligibility determination is supposed to provide more than an outcome. It is supposed to expose enough of the basis for the affected person to understand and challenge what happened.

If an AI system generates a plausible rationale after the fact, it may satisfy the institution’s appetite for procedural form while weakening the affected person’s ability to contest the real basis of action.

The result is not opacity in the simple sense. It is worse: the appearance of transparency generated by a system that may be concealing the actual decision chain.

Two obvious swamps

Benefits administration is the cleanest case because the stakes are material and the asymmetry is brutal. Eligibility systems already compress messy lives into categories, proofs, deadlines, and exceptions. AI promises speed: summarize files, identify discrepancies, detect fraud, recommend approval or denial, generate notices. The crucial distinction is whether AI remains bounded to a role. A model that helps identify missing documents is one thing. A model whose discrepancy flag becomes the practical reason for denial is another. A generated notice that accurately reflects a human decision is one thing. A generated notice that manufactures individualized reasoning around a workflow default is another.

Content moderation is the other obvious swamp because it has the strange combination of high consequence and theatrical reversibility. An account suspension is not imprisonment, but it can still affect livelihood, speech, reputation, organizing capacity, and access to social infrastructure. Platforms love the language of reversibility: appeal, review, restore, escalate. But reversibility after damage is not the same as admissibility before action. A label affects visibility. Reduced visibility changes behavior. Behavior becomes data. Data confirms the label. Future systems inherit the prior moderation residue as context. The user can appeal the suspension, maybe. They cannot usually contest the compression ecology that made them administratively legible as the kind of account that should be suspended.

Both cases show the same pattern: the output is not merely wrong or right. The output is trying to stand somewhere inside a decision chain.

That standing must be governed.

The rule is verbs, not vibes

The relevant governance question is not whether AI may participate in administration. It already does, and it will do more.

The question is which administrative verbs it is allowed to perform.

Can it observe? Under what provenance requirements?

Can it interpret? Under whose categories?

Can it recommend? With what uncertainty and review burden?

Can it authorize? Usually no, unless a very specific standing relation has been established.

Can it act? Only where consequence is bounded, reversible, and institutionally owned.

Can it generate receipts? Yes, but receipts do not legitimate what they record.

Can its outputs persist? Only with decay, provenance, revalidation, and contestability.

This is less glamorous than asking whether AI is intelligent. It is also much closer to where the damage will happen.

A model output is not an administrative fact until an institution admits it as one. A recommendation is not an authorization. A receipt is not a justification. A proxy does not measure the target merely because the institution acts through it. A durable record does not carry fresh admissibility merely because it persists.

AI administration will not produce post-bureaucratic governance. It will produce hybrids of bureaucracy, statistical compression, human review, vendor infrastructure, managerial pressure, and political command.

Some hybrids will be useful. Some will be dangerous. Most will be useful in exactly the way that makes them dangerous.

The task is not to preserve bureaucracy as it exists. Existing bureaucracies already compress badly, hide discretion, displace responsibility, and exhaust the people subject to them. But replacing visible institutional mediation with less legible machine-mediated compression does not solve those failures. It risks making them harder to contest.

AI does not eliminate bureaucracy’s tradeoffs. It relocates them into a less legible mediation layer. Once that layer becomes the place where decisions are sensed, authorized, acted on, and remembered, the question is no longer whether the machine is smart.

The issue is not whether systems include provenance, oversight, or context. The issue is whether those artifacts can bind authorization.

The question is what the institution has allowed the machine’s compression to become.