When Political Concepts Become Machine-Readable

Imagine a public agency assessing whether it retains sufficient control over an AI service. Its assessment system checks the contract, confirms that the agency can revoke the provider’s access, and verifies that the shutdown mechanism works. It reports that control is preserved.

Yet the service now supports several essential administrative functions. The agency has retired its previous workflow, reassigned the people who understood it, and stopped maintaining an independent copy of the operational data. Exercising its contractual rights would interrupt services beyond an acceptable period.

The assessment may have applied its criteria correctly. Those criteria nevertheless captured only part of what control requires. The immediate problem is incomplete operationalization: formal authority and a functioning shutdown mechanism have been treated as sufficient evidence of control. The example does not by itself establish a disagreement over the meaning of control.

A different problem arises when the concept itself is contested. Competing accounts of legitimacy, autonomy, or trust may disagree about what should count, even when the available evidence is complete. Better measurement cannot settle that disagreement.

These problems call for different responses. Incomplete operationalization calls for clearer evidence requirements and checks that the measures adequately represent the concept. Conceptual disagreement requires an explanation of the interpretation selected, authority for its institutional use, and procedures for challenge and revision. Both problems can occur in the same assessment.

This is a political problem embedded in a technical decision: which meaning of control has become operational, who selected it, and what consequences follow from that choice?

As AI participates in administrative work, questions of this kind become part of system design. A system may classify situations, rank risks, interpret instructions, recommend interventions, or help determine which cases receive attention. Each activity requires distinctions. Some are relatively straightforward. Others concern authority, dependence, discretion, responsibility, or a person’s practical ability to challenge a decision.

When such distinctions influence institutional action, their definitions become part of the machinery of governance.

The underlying problem predates AI. Institutions have long translated complex concepts into forms, eligibility rules, indicators, and decision procedures. Research on computational fairness has examined how an operational measure can diverge from the concept it claims to represent. Jacobs and Wallach’s Measurement and Fairness, for example, distinguishes disagreements about measurement from disagreements about the underlying concept itself. Measurement and Fairness.

AI gives this problem another setting. A particular interpretation can be reproduced across many decisions, including through systems whose treatment of ambiguous cases is difficult to inspect. The place where a consequential definition is selected can also shift. Legislation, regulation, policy, and administrative procedure provide identifiable sites for authorization and challenge, however imperfect those arrangements may be. Substantive choices can instead be made in a prompt, dataset, rubric, evaluation framework, product requirement, or software implementation, without an equivalent process for authorizing and contesting them. Unless that shift is made explicit, an implementation choice can quietly acquire institutional authority.

Consider a few apparently simple propositions.

A person continues using a service. Does that establish trust, or could it reflect the absence of a practical alternative?

An institution has the right to intervene. Does it also have the information, resources, and available means to exercise that right?

A system follows instructions. Does that establish agreement about goals, or only compliance in the observed circumstances?

A complaint mechanism exists. Can the affected person actually use it, and can a successful complaint produce an effective remedy?

Each question separates properties that ordinary language often gathers under one word. A machine-assisted procedure needs a defensible way to handle those distinctions.

This creates three connected tasks.

The first is conceptual clarification: specifying what a term means in the particular context, which distinctions matter, and which interpretations remain contested.

The second is operationalization: identifying the evidence that supports a classification. If an institution is described as dependent on a provider, the assessment should specify the function, the relevant alternatives, the time horizon, and the consequences of losing access.

The third is preserving meaning through implementation: checking whether the system’s criteria retain the distinctions and qualifications that justified the concept in the first place.

Technical abstraction can make this difficult. A model may represent the parts of a situation that are easiest to encode while excluding the relationships that make the classification politically significant. Selbst and colleagues describe this problem in their analysis of abstraction in sociotechnical systems. Fairness and Abstraction in Sociotechnical Systems.

A fourth task follows: governing the formalization itself.

Who may authorize a definition for a particular institutional purpose? Who can challenge its application? Who can argue that the definition is inadequate? What happens to earlier decisions when the criteria change?

Political researchers can clarify concepts and expose assumptions. Engineers can make implementations inspectable and test their behavior. Public authorities must account for the standards they authorize. Affected people need meaningful opportunities to contest classifications and their consequences. None of these roles can simply substitute for the others.

Nor should formalization promise to eliminate disagreement. Different understandings of legitimacy or autonomy may express substantive political differences. A responsible system must sometimes preserve those differences, identify insufficient evidence, or refer a question to an authorized decision process. Producing a classification in every case is not a sufficient measure of success.

At Indispensable Systems, we are developing a research track around these problems. Its starting point is the relationship between actors and the architectures through which their ability to act is supported, constrained, and redistributed.

Our particular concern is what happens when necessary functions become dependent on systems that are difficult to replace. Such dependence can change the practical meaning of control, delegation, and exit even while contracts and technical permissions remain intact.

The intended contribution is an explicit and contestable analytical grammar: distinctions, evidence requirements, and review procedures that people and AI systems can use together. It should make classifications, their grounds, and disagreements about them visible, so that people can ask whether a classification is supported, whether the chosen concept fits the situation, and whether its institutional use is justified.

Our vocabulary makes no claim to provide a neutral or uniquely correct account of political reality. This applies to our own distinctions between agency and autonomy, or dependence, reliance, and trust. They are analytical proposals with a scope and rationale that must remain open to challenge. Even the choice of which distinctions matter is not neutral; our vocabulary must be subject to the same scrutiny we propose for other institutional classifications.

The conceptual infrastructure must itself remain open to revision. Once definitions are embedded in procurement requirements, data structures, and audit procedures, changing them can become difficult. Dependence on the system may also make its classifications difficult to revise: records, workflows, and assessments may all presuppose the same distinctions. This is a possible form of conceptual lock-in. Indispensability may attach not only to infrastructure, but also to the conceptual classifications embedded in it. Their provenance, scope, uncertainties, and amendment procedures therefore belong in the design from the beginning.

The public need is concrete: people must be able to understand, challenge, and revise consequential institutional judgments—including the definitions through which those judgments are made.

The companion article, More Capable, Less Able to Change Course?, works through these distinctions in the assessment of control: what follows when formal authority, technical functionality, practical feasibility, and evidence of actual exercise are examined separately?