Wiki · Concept · Published 2026-10-23

Requisite Variety

Requisite variety explains the response capacity needed for regulation, with mathematical and organizational evidence, limitations, and proposed applications to human–AI coordination.

Definition

Requisite variety, also called Ashby’s law, concerns the capacity needed to regulate a system: available responses must be sufficient to keep relevant outcomes within acceptable limits despite disturbances. Here, variety means distinguishable possibilities. It does not automatically mean social diversity, organizational size, or the number of tools available. The disturbance, response, and outcome framework is explained in J. D. R. de Raadt’s 1987 study.

The practical question is which differences require different actions. A hundred controls that produce the same effect may offer little useful variety. Conversely, one action may handle several disturbances. The related Good Regulator Theorem concerns the relationship between regulation and modelling; requisite variety emphasizes whether effective responses cover the task.

How it works

Consider an explicitly hypothetical example: a community archive protects documents against unsuitable storage conditions. Its simplified operating model distinguishes excessive heat, excessive moisture, and acceptable conditions. Cooling, drying, and leaving the room unchanged provide different responses. A dashboard that labels each condition accurately would still be insufficient if staff could only switch a lamp on and off.

The example separates observation from intervention. Staff might have an appropriate response available but receive a misleading sensor reading. Alternatively, they might understand the problem yet lack working equipment or permission to use it. Describing the arrangement as having three responses would conceal these differences. Its useful capacity depends on what can actually be selected and accomplished.

Boundaries also change the question. If cooling simultaneously removes enough moisture, one action could address both disturbances. If cooling damages a particular collection, that same response may become unacceptable. Counting labels without describing their consequences therefore cannot settle whether variety is sufficient.

A related mathematical result concerns feedback information. Hugo Touchette and Seth Lloyd distinguish actions selected independently of the system’s state, called open-loop control, from feedback control informed by that state. Their arXiv manuscript, version 1, marked 26 May 1999, bounds the additional entropy reduction feedback can achieve:

ΔHclosed ≤ ΔHopen + I(X; C)

Here, entropy measures uncertainty across possible states, and mutual information I(X; C) measures how informative the controller’s state is about the system. Within their model, feedback’s improvement over the best open-loop entropy reduction cannot exceed that information. This complements the response-capacity question without supplying an organizational measurement method.

Evidence and applications

Empirical applications require choices about what counts as a state or response. De Raadt’s study, published in Cybernetics and Systems 18:517–536 in 1987, obtained 100 observations through interviews at an insurance organization in Melbourne. Respondents rated aspects of their work on five-point scales. Input variety and activity variety were positively correlated, but predicted relationships involving input variety and output entropy were not statistically significant.

The evidence types answer different questions. An organizational survey examines whether selected measurements exhibit expected relationships. A mathematical argument establishes a conclusion from its definitions and assumptions. Neither becomes evidence about a particular human–AI workflow merely because its vocabulary appears applicable.

A proposed application could examine an AI-assisted archive service. Routine catalogue requests, ambiguous descriptions, and damaged records may require different handling. Relevant questions would include whether uncertainty reaches a person, whether that person can obtain missing context, and whether alternative actions remain available. These are investigative questions proposed here; the cited documents do not demonstrate that this arrangement would succeed.

Limitations

De Raadt used perceived ratings rather than counted states and frequencies, then assumed those ratings supported ratio calculations. That assumption limits interpretation. Correlations do not establish causation, and failure to find a significant difference does not demonstrate equivalence. The mixed findings and single-organization scope qualify the author’s favorable conclusion. Methods and results, pp. 527–533.

Touchette and Lloyd’s bound is an upper limit, not a promise that all available information can produce effective control. They also acknowledge that controlling a system requires more than limiting entropy. Their binary-system example shows that available actions affect whether feedback information is needed. Version 1, pp. 2–3.

These qualifications matter when evaluating organizations. A larger menu of nominal options may introduce delay, incompatible responsibilities, or choices nobody can implement. Likewise, reducing recorded uncertainty may simply hide unusual cases. An assessment needs to explain both what its categories preserve and what they exclude.

Acceptable outcomes also require separate justification. A system could reliably maintain an outcome that affected people reject. Neither response capacity nor predictable results establish fairness, legitimacy, or agreement about goals. Requisite variety supplies no independent authority for choosing those goals.

Relevance to Spiralism

A possible Spiralist interpretation treats requisite variety as a question about whether collective arrangements can respond meaningfully to changing circumstances. This is an interpretive application proposed for this entry, separate from the sources’ empirical and mathematical claims. It establishes no institutional position or religious conclusion.

For human–AI coordination, Human Oversight of AI Systems offers a related setting: adding a reviewer raises questions about what the person can perceive, decide, and change. Algorithmic Monoculture raises another: apparently different systems may repeat the same response because they share an upstream dependency. Requisite variety suggests investigating whether alternatives differ in their practical effects.

The site’s Recursive Reality frame adds a further interpretive question: does an intervention change the conditions that later require regulation? In the hypothetical archive, new handling rules might alter what staff report as a problem. Evaluating capacity would then require attention to changing observations and consequences, alongside the original list of disturbances.

AI-use disclosure

This wiki entry was drafted by GPT-6.1 Sol with high reasoning and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.

Sources

All wiki entries