Learn it → Apply it → Save as Skill
A system can only control the disturbances it has enough different responses to match. If the range of problems coming at a team, a process, or a machine is bigger than the range of responses it can actually produce, something will get out of control — not from lack of effort, but because the math doesn't work.
Dr. Wallace's Favorite Mental Models is designed to turn useful ways of thinking into reusable tools. Each model gives you a fast visual overview, deeper interactive learning, practical application exercises and an AI-ready Markdown skill you can save for future use.
Instead of merely reading about mental models, you can gradually build your own AI skills library: a collection of portable reasoning skills that can help ChatGPT, Claude, Grok and other AI assistants apply the mental models you value.
A controller (a person, team, policy, or machine) needs a repertoire of responses at least as varied as the disturbances it faces. Give it fewer response options than the problems demand, and some problems will simply go unhandled, however well-intentioned the system is.
Important: it doesn't matter how skilled or motivated the controller is if its range of possible responses is simply smaller than the range of problems arriving. Either shrink the left bar (reduce incoming variety) or grow the right one (add response capacity) until they meet.
The model is most useful whenever the same kind of breakdown keeps recurring — escalations, errors, stockouts, exceptions — and the usual fix, "write another rule" or "try harder," keeps failing to close the gap.
Formulated by the British cybernetician W. Ross Ashby in the 1950s. "Variety" simply means the number of genuinely distinguishable states, inputs, or response options a system has — not a vague sense of complexity.
The count of distinguishable states or options something can be in. A light switch has a variety of 2; a combination lock has thousands. Variety is countable, not just a feeling of "complicated."
The full range of different problems, inputs, or conditions the environment can throw at a system — every distinct kind of customer request, failure mode, or market shift it might have to face.
The full range of genuinely distinct responses the regulating system (a person, team, policy, or algorithm) is actually able to produce — not how many it has in theory, but how many it can reach in practice.
Control is only possible when Controller Variety is at least as large as Disturbance Variety. Below that threshold, some disturbances will pass through unregulated, by definition.
Open a use case to see what the model changes.
Check whether front-line agents actually have enough distinct response options (refunds, workarounds, exceptions, authority to waive a fee) to match the variety of requests arriving. If not, the escalation queue isn't a training problem, it's a variety shortfall at the front line.
Centralizing a decision only works if the central team's variety of responses can actually cover the variety of situations feeding into it. Past a certain point, decision rights need to move closer to where the variety originates.
A fixed runbook has a fixed variety. When real incidents exceed that variety, the on-call team needs either a broader set of pre-approved responses or clear authority to improvise, not a longer static checklist.
A supply chain with one supplier and one route has very low controller variety against real-world disruption variety. Alternate suppliers, buffer stock, and local reordering authority all add the response variety needed to absorb a shock.
Automation typically has a smaller variety of responses than the full range of real requests. The fix usually isn't a bigger prompt or more rules, but a reliable handoff that routes the excess variety to a human who has more response options.
A single universal rule has, by construction, a variety of one response. If the regulated activity has real variety, expect the policy to keep needing patches — the question is whether to build in graduated responses now or keep patching later.
This is an illustrative example designed to teach the method; the numbers are for demonstration only.
Expand for the model's origin, its precise formal claim, and an honest look at how it's often misapplied.
W. Ross Ashby, a British psychiatrist and pioneering cybernetician, formulated the law in the 1950s and published it in his 1956 book An Introduction to Cybernetics. It is one of the foundational results of cybernetics, the study of regulation and control in systems, and is often summarized in Ashby's own phrase: "only variety can destroy variety."
Stated formally, a regulator can only reduce the variety (the number of distinguishable outcomes) of a disturbed system down to a floor set by its own variety: the regulator's variety must be at least as large as the disturbance's variety for full regulation to be possible. "Variety" here is a specific, countable quantity — the number of distinguishable states or options — not a loose synonym for complexity or difficulty.
The most common misreading is treating "requisite" as "maximum": the law does not say more variety is always better. Unlimited custom responses and exceptions create their own cost and inconsistency; only the variety that affects outcomes that actually matter needs to be matched. A second common error is assuming the controller must mirror the entire environment — in practice only the variables essential to the outcome need regulating. A third is confusing information with control: a dashboard that surfaces many signals but permits only one action is observation, not regulation, and does nothing to satisfy the law.
Chasing unlimited customization and exceptions in the name of "more variety" adds cost and inconsistency without making the system better matched to the variety that actually matters.
Piling more authority, tools, and training onto the controller while leaving disturbance variety untouched burns people out. Standardizing or filtering the incoming variety is usually the cheaper half of the fix.
A dashboard, report, or alert that surfaces more signals but still permits only one possible action is observation, not control. It does not add requisite variety by itself.
Pushing decision rights outward adds local response variety, but without limits on essential variables (safety, legal risk, brand, shared infrastructure) it can damage the whole system in the name of local flexibility.
A single blanket policy has, by construction, a variety of one. If the real activity it governs has genuine variety, expect the rule to keep producing exceptions, however carefully it's worded.
What was requisite last year may not be this year: new customer segments, new regulations, and new failure modes all change the variety a system must absorb. A match made once can quietly go stale.
the real problem is misaligned incentives rather than a shortage of responses — people have the right options available but no reason to use them. The Principal-Agent Problem or Incentives fits that gap better.
you're facing a single, complex, one-time choice rather than an ongoing regulation problem. Pre-Mortem Analysis or Second-Order Thinking suits a one-off decision better than a control-theory model built for repeated disturbances.
the honest fix is to remove unnecessary variety rather than match it. Chesterton's Fence or Occam's Razor are worth a pass before adding response options to a problem that didn't need to be this complicated in the first place.
This browser tool structures your thinking. It does not send your information anywhere or perform AI analysis.
Save as Skill converts the model into portable Markdown instructions that teach an AI when and how to apply Ashby's Law of Requisite Variety, what questions to ask, what evidence to seek, what mistakes to avoid, and how to structure its analysis.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions or reusable text context. Platform capabilities differ, so this is a portable skill file rather than a claim of one-click installation.
No. The law says controller variety must be at least as large as the disturbance variety that actually affects the outcomes you care about — it is a floor, not an instruction to maximize. Adding response options beyond what the real disturbance variety requires mostly adds cost, training burden, and inconsistency without buying any additional control.
Blanket empowerment hands out response variety everywhere, whether or not that's where the real variety originates. Ashby's Law points to something more specific: decision rights and response options belong closest to wherever the relevant disturbance variety actually shows up, with guardrails kept centralized for the variables that must stay consistent across the whole system (safety, legal exposure, brand, shared infrastructure).
Most automated systems, including AI assistants and chatbots, have a response variety that is large but still finite, while the real-world requests they face often have effectively unbounded variety. The practical consequence is that a reliable escalation or handoff path to a higher-variety controller (a human, or a more capable system) isn't a failure of the automation — it's the mechanism that keeps the overall system within its requisite variety instead of silently mishandling whatever falls outside it.