In 1956, a British psychiatrist named W. Ross Ashby published An Introduction to Cybernetics, a book that would become foundational to fields as disparate as artificial intelligence, ecology, management science, and systems biology. Ashby was interested in a deceptively simple question: how does anything control anything else? How does a thermostat regulate temperature? How does an organism maintain homeostasis? How does any system impose order on another system that would otherwise drift into chaos?

His answer was the Law of Requisite Variety. Buried in Chapter 11, amid discussions of game theory and information channels, Ashby made an observation that is a core axiom of instructional design.

“If R’s move is unvarying, so that he produces the same move whatever D’s move, then the variety in the outcomes will be as large as the variety in D’s moves. D now is, as it were, exerting full control over the outcomes.”

Put simply, Ashby was describing a game between two players: R, (the regulator) trying to achieve a particular outcome, and D, (the source of disturbance). The regulator wants to keep the outcome within acceptable bounds. The disturbance wants to push it outside them. Who wins depends on one thing: whether R has enough different responses to counter D’s different moves.

This condition applies wherever one system attempts to shape the behaviour of another. Teaching and instructional design, at its core, is an attempt to stabilise particular cognitive outcomes in the face of learner variation. It is in essence, a regulatory enterprise.

Consider a child learning to read in a reception classroom. The teacher is R: the regulator, trying to ensure the child learns to decode words by mapping graphemes to phonemes and blending them fluently. The child’s repertoire of strategies is D: the source of variation. The child has a range of strategies; they can sound out the letters properly, or guess from the picture, or memorise the shape of the word, or rely on the first letter and fill in the rest from context, or simply repeat what a classmate has just said. Five strategies, perhaps more.

The teacher asks the child to read a word from a simple book. The word is horse. The child says “horse.” The teacher hears the correct response and moves on. One response. Correct or incorrect.

Ashby’s law tells us what follows. If the regulator’s move is unvarying, or to put it more simply, if the only thing being checked is whether the word sounded right, then the variety in the child’s strategies determines the variety in the outcomes. If the child guessed from the illustration, or from the first letter and context, that strategic variety has passed through unblocked. The outcome variable is satisfied. The internal process is unregulated.

In that moment, control has shifted. The teacher believes reading has occurred. The system records success. But the child may not have decoded anything at all. The appearance of learning is preserved, however the underlying cognitive operation is not.

The law can be stated in a single sentence: to control a system, you must have at least as many possible responses as the system has ways of going wrong. Or as Ashby put it even more tersely: “only variety can destroy variety.”

In other words, if learners have more ways to succeed than your system has ways to detect and block easy shortcuts, the shortcuts will win. Not because learners are lazy or devious, but because they are rational agents navigating a system that has left paths of least resistance wide open. Learning will flow toward whatever strategy produces the right answer with minimal effort. This is not a character flaw in students; it is a predictable consequence of inadequate regulatory variety. The fault lies not in the student but in the design.

Why Differentiation Fails

One strategy that attempted to address low-variety regulation, or the tendency to check only whether answers are correct rather than how they were reached, in conventional classrooms was differentiation. Faced with obvious variation between pupils, the instinct was to increase variation on the instructional side: different tasks for different groups, scaffolded worksheets, extension problems, adjusted levels of difficulty. In principle, this appears to expand the teacher’s repertoire of responses.

In practice, it often failed because it mistook variation in performance for variation in strategy. It multiplied tasks without increasing the system’s capacity to detect how those tasks were being solved.

Ashby’s law is not about surface diversity. It is about whether the instructional system’s variety matches the disturbance that actually matters. If the disturbance lies in strategy, in how the child arrives at the answer, then varying the difficulty of the text does little to address it. A child who guesses from pictures will continue to guess from pictures on an easier book and also on a harder one. A pupil who has learned to pattern-match in mathematics will pattern-match on the basic questions and on the extension ones.

The deeper problem is informational. An instructional system can only regulate what it can detect and many learning environments rely on a channel of extremely low capacity: correct or incorrect.

Correct or incorrect carries almost no information about process. It does not distinguish decoding from guessing, understanding from memorisation, reasoning from elimination. If the system cannot detect how success was achieved, it cannot constrain how success was achieved. The range of strategies available to the learner, what Ashby calls “the disturbance space”, remains uncontrolled.

This is why so much classroom performance is fragile. Students appear fluent in the moment, yet falter when the context shifts. They were “regulated” at the level of outcome, not at the level of process. The system ensured that answers were right often enough, but it never ensured that the right thinking had occurred. The system is, in Ashby’s terms, informationally impoverished; and no amount of pedagogical enthusiasm can compensate for what the channel cannot carry. In other words, you can’t out-teach a system that cannot see.

When Correct Answers Conceal Cognitive Failure

I’ve written before about the problem of students getting answers “correct” but learning very little. From a cybernetics perspective, a binary answer of correct or incorrect is a channel with low capacity. It tells us little about how an answer was produced. And if a system cannot detect how success was achieved, it cannot constrain how success was achieved. In Ashby’s terms, the disturbance retains control.

This does not mean binary assessment is inherently flawed. In some cases, it is entirely sufficient. If the task tightly aligns with the target cognitive operation, and there are few viable shortcuts, then correct or incorrect may be all the regulation required. A simple factual retrieval question asked in open response form leaves little room for alternative strategies. The disturbance space is small. The regulator has enough variety. Right or wrong might also be a very good way of building knowledge and diagnosing misconceptions early.

The problem arises when the disturbance space is large. By disturbance space, I mean the range of different ways a learner can approach a task while still arriving at an acceptable answer. It is the full repertoire of strategies, shortcuts, heuristics and workarounds available in that moment.

When multiple cognitive strategies can produce the same correct answer, binary scoring becomes blind. Recognition can masquerade as recall. Elimination can masquerade as understanding. Pattern familiarity can masquerade as reasoning. The system records success, but it has no way of knowing what produced it.

Consider a pupil revising the causes of the First World War. The retrieval question reads:

Which of the following was a long-term cause of the war?
A. The assassination of Archduke Franz Ferdinand
B. Militarism
C. The sinking of the Lusitania
D. The Treaty of Versailles

The pupil selects B. But what just happened cognitively? Perhaps the pupil genuinely retrieved a schema in which militarism, alliances, imperialism and nationalism form an interconnected explanation of escalating tensions across Europe. That is one possibility.

Or perhaps the pupil simply remembered that the assassination was a trigger event, not a long-term cause. Two options can be eliminated immediately. The Treaty of Versailles obviously came after the war. The sinking of the Lusitania happened later in the conflict. “Militarism” is the only one that sounds plausibly abstract and structural. So B must be right.

The correct answer is the same. The cognitive operations are not. In one case, knowledge has been retrieved and integrated. In the other, superficial cues have been exploited. Binary scoring cannot distinguish between them. The regulator sees success. The strategic variety that produced it remains invisible.

This in part explains a fundamental problem in formal schooling; students to perform well on class assessments over the year and then underperform in an exam. It is tempting to attribute this to nerves, poor revision habits, or exam technique, and sometimes that is true. But often something more structural is at work.

During the year, the disturbance space was constrained. Question types were familiar. Contexts were predictable. Teachers gave hints, prompts and scaffolds. Retrieval was supported by cues embedded in the classroom environment. In effect, the regulator tolerated a degree of strategic shortcutting because the surface behaviour looked acceptable.

Ashby’s law does not condemn binary scoring. It simply reminds us that regulation requires sufficient variety. When the disturbance space grows and the regulator remains simple, control shifts. The learner adapts. The system records success. Learning, meanwhile, may remain fragile.

The Promise and Peril of Adaptive Systems

This is where adaptive learning platforms and AI tutors become genuinely interesting, and where Ashby’s framework becomes most revealing about their limitations.

In principle, these technologies offer something that traditional instruction cannot: a dramatic expansion of regulatory variety. An AI tutor can observe not just whether the answer is correct, but how long the learner took, what errors they made along the way, whether they asked for hints, how they responded to follow-up probing, whether their pattern of success suggests genuine understanding or strategic elimination.

Where a paper test carries one bit of information per problem, correct or incorrect, an intelligent system can potentially harvest dozens of signals per interaction. Where a human teacher monitoring thirty students can respond to perhaps a handful of strategic variations per lesson, an adaptive system can in principle track each learner’s idiosyncratic patterns continuously.

That is regulatory variety of a genuinely different order. The channel widens. More of what actually matters becomes, at least in principle, visible.

But increased variety in principle is not the same as requisite variety in practice, and this is where the current generation of adaptive platforms reveals its limits. Most adapt along a single dimension: difficulty. If the learner performs well, the system serves harder questions; if performance drops, it serves easier ones. This is regulation of a kind, but of a peculiarly narrow kind. It addresses one dimension of variation, the learner’s current performance level, while leaving strategy entirely unregulated.

A learner who has discovered that process of elimination works reliably on multiple-choice questions will sail through such a system. The platform will interpret sustained correct answers as mastery and serve progressively harder material. The learner will apply the same shortcut to those as well. The system will report that learning has occurred. The learner will have acquired nothing except the knowledge that process of elimination works. The regulator has variety; it simply does not have the variety that matters.

This is the central disappointment of much educational technology: not that it lacks sophistication, but that its sophistication is deployed in the wrong direction. Engagement metrics, gamification, personalised difficulty curves, these are genuine technical achievements. They are also, from Ashby’s perspective, regulatory activity aimed at the wrong variable. A system that optimises for time-on-task while leaving cognitive strategy unmonitored is not solving the problem; it is measuring a proxy for it and mistaking the proxy for the thing itself.

Designing for Requisite Variety

Taking Ashby’s law seriously means treating instructional design as an engineering problem rather than an art, and that shift in framing has real consequences for how we build and evaluate educational systems.

It begins with a question that many educators do not ask systematically: how else could a learner succeed here without doing the intended thinking? Every task has alternative solution paths, and those paths will be found. The designer’s job is to map them and close them, not by making tasks harder, but by making shortcuts unreliable. A task well-designed against Ashby’s law is one in which the target cognitive operation is not the easiest path to the correct answer; it is the only path. This is an instructional invariant; a non-negotiable condition of effective design, as fixed and indifferent to our preferences as any law in engineering.

All this requires a somewhat different conception of assessment. The question is not “did they get it right?” but “what would they have to have done cognitively to get it right in this way?” Diagnostic power matters as much as correctness. Requiring worked solutions, varying problem formats so that pattern-matching fails, asking for explanations that cannot be generated without understanding, probing with follow-up questions that are only answerable if the first answer was genuinely reasoned rather than guessed: these are not merely good pedagogical habits. They are mechanisms for expanding the instructional system’s regulatory variety to match the strategic variety of learners.

For adaptive systems and AI tutors, the implication is architectural rather than cosmetic. Adjusting difficulty is insufficient if the learner’s strategy works across difficulty levels. The system must adapt to strategy, not merely to performance: detecting when a learner is pattern-matching and breaking the pattern, detecting when a learner is guessing and making guessing unreliable, detecting when a correct answer was produced by a process that will not transfer and probing until that becomes apparent. This is not a matter of prompting an AI more cleverly. It is a matter of building diagnostic and constraint mechanisms into the system’s design from the outset.

And then, finally, there is the upper bound that no amount of ingenuity can dissolve. No system can regulate what it cannot detect. Learner ingenuity will always exceed designer foresight; there will always be shortcuts that were not anticipated, strategies that were not mapped, paths that were left open by accident. Requisite variety is an asymptote, not a destination. The honest response to this is not despair but iteration: observe what learners are actually doing, diagnose which strategies are succeeding without learning, close those paths, and observe again.

The Upper Bound: What the System Cannot See

Ashby was not thinking about classrooms. He was thinking about thermostats, organisms, and the abstract mathematics of control. But I find this engineering mindset fascinating when applied to learning and I think his insight cuts across domains because it concerns the logic of regulation itself, and that logic applies wherever one system attempts to constrain another.

Instructional design is, at its core, an attempt at regulation. We want particular knowledge structures to form and consolidate; we want others not to. The learner arrives with immense strategic variety, countless ways of navigating any task we set, many of which lead to the appearance of success without the substance of it. The instructional system is the regulator: its job is to match that variety, to close the paths that lead to performance without learning, to ensure that the only reliable route through the task is the cognitive operation we actually intend.

If the system cannot do this, the learner’s variety determines the outcome. The platform reports mastery. The assessment records a pass. Learning, in any meaningful sense, may not have occurred.

This is not a counsel of despair about teachers, or technology, or learners. It is a precise description of what instructional design must accomplish and why it so often falls short. The problem is not effort or intention; it is regulatory variety. The system checks too little, detects too little, and consequently controls too little. Learners, being rational, take the paths that are left open.

“Only variety can destroy variety.”