Systems & Complexity

Why Good Reforms Keep Failing

A new curriculum is mandated, and teachers adapt it in ways its designers never anticipated. Technology is deployed to “personalize learning,” and it reinforces the inequities it was meant to address. Assessment is redesigned, and schools learn to satisfy the metric without improving what the metric was meant to measure.

We would locate the common cause not in any of these interventions but in an assumption underneath all of them: that education is mechanical—something repaired by replacing broken parts or adjusting isolated variables. Install new software. Hire more staff. Mandate different tests. Underneath sits a commitment to linear causality: change X, and Y follows.

But education is a complex adaptive system—a living network of people, institutions, technologies, and relationships interacting over time, producing outcomes no single actor controls. Cause and effect are rarely proportionate. Small changes cascade. Well-designed interventions backfire. Solutions that work beautifully in one district fail completely in the next.

This is not an argument for fatalism, and we would not want to be read that way. Complexity is not an excuse for inaction; it is a description of the conditions under which action has to be designed.

The Engines: Feedback Loops

If you want to know why a school system behaves as it does, look for its loops. They are the engines—the reason conditions improve, deteriorate, or stubbornly hold still. Two examples, both of them ordinary:

Nothing distinguishes these two loops structurally. The same mechanism produces both. And AI enters directly into that mechanism: tutoring systems that give timely, actionable feedback can accelerate the first, while algorithmic sorting of students into tracks can entrench the second. Speed without judgment amplifies whichever loop it finds.

Where to Push

Not all interventions are equal. Drawing on Donella Meadows’ hierarchy of leverage points, we distinguish the changes that are easy to make and change little from the changes that are hard to make and change everything.

Most educational technology, including most AI, is sold at the top of that left column and justified with the language of the right. Our research maps where AI genuinely strengthens high-leverage points—and where it entrenches low-leverage thinking at unprecedented scale.

What AI Actually Does to a System

AI does not simply add capability. It restructures the dynamics, often subtly and on a delay.

It compresses some feedback loops and hides others, giving real-time signal on student understanding while creating the illusion of instant impact in a process that unfolds over years. It creates new information flows that can surface inequity earlier—and simultaneously redistribute power, eroding professional autonomy when administrators gain real-time visibility into teacher practice. It shifts stocks and flows rather than eliminating them: teachers spend less time grading and more time learning platforms, interpreting recommendations, and troubleshooting failures.

Whether any of this strengthens education depends less on the quality of the tool than on how well the system receiving it is understood.

How We Work

We map before we intervene. Who are the actors? What loops already operate? What delays obscure cause from effect? This mapping routinely reveals that the problem AI was bought to solve is a symptom of something else—a district seeking AI grading to cut teacher workload, when workload is driven by class sizes and reporting requirements.

We track over years, not months. Pilot studies capture initial reactions. They are structurally incapable of seeing the trade-offs that matter: the system that raises short-term scores while undermining deeper learning, the tool that gains efficiency while eroding the relationships motivation depends on. We regard the short evaluation window as one of the more consequential methodological weaknesses in this field, including in work we admire.

We design for adaptation, not compliance. Educational systems will adapt around whatever is introduced. So we favor modular architectures, local configurability, transparent logic that can be inspected and challenged, and exit strategies that permit graceful disengagement rather than lock-in.

We treat equity as a system property. From a systems view, inequity is not an unfortunate side effect. It is a stable pattern held in place by reinforcing loops—well-resourced schools attract experienced teachers, outcomes improve, property values rise, more resources flow in. Changing that requires changing structure, not improving components.

The Path Forward

Education will not be “solved” by AI or any other technology. It can be strengthened—made more equitable, adaptive, and humane—if we approach it with systems literacy: mapping loops before deploying at scale, finding the leverage points where AI can genuinely catalyze change, attending to the delays that shape long-term outcomes, and building educators’ capacity to navigate complexity rather than imposing rigid solutions on them.

Education is not a machine to be fixed. It is an ecology to be tended. Tending it well begins with seeing it clearly—in its complexity, its interdependence, and its genuine possibility.