Feedback Loops

Definition

A feedback loop exists whenever the output of a system circles back to influence its own input — effects become causes, and the system actively shapes its own future behavior.

There are two fundamental kinds. A reinforcing loop amplifies whatever direction the system is already moving: growth feeds more growth, reputation begets more reputation, a bank run accelerates as each withdrawal makes the next more likely. A balancing loop counteracts change, pushing the system back toward a target or range: a thermostat, a body temperature regulation mechanism, or the pricing mechanism in a competitive market. Most real systems contain multiple reinforcing and balancing loops operating simultaneously, often with significant delays between cause and visible effect. The interplay among these loops — and the timing of those delays — is what makes complex systems so routinely surprising.

Why it matters

How it works

Reinforcing loops: the engine of compounding and collapse

A reinforcing loop has no inherent direction — it amplifies whatever the system is already doing, whether that is good or bad. On the upside, this is the mechanism behind compounding interest, network effects, skill acquisition, brand reputation, and viral content. Each unit of output generates slightly more input than it consumed, so the next cycle is bigger. The pattern feels slow and then suddenly overwhelming, which is why people consistently underestimate compounding in the short run and overestimate it in the long.

On the downside, the same structure drives bank runs, addiction spirals, ecological collapse, and social conflict. A company loses a few key employees; the workload on remaining staff increases; morale declines; more employees leave; the burden on those remaining grows further. The loop is identical in structure to a success spiral — only the direction differs. Recognizing a reinforcing loop does not tell you which direction it is running; it tells you that whatever direction it is running, it will keep accelerating until a balancing loop intervenes or a resource constraint is hit.

Balancing loops: the machinery of stability

Balancing loops are goal-seeking: they hold a system variable near a target by generating corrective action whenever actual state diverges from desired state. The thermostat is the textbook example, but the pattern is everywhere. Central banks adjust interest rates to hold inflation near a target. The immune system mounts a response when pathogens exceed a threshold and stands down when they are cleared. Competitive markets lower prices when supply exceeds demand and raise them when demand exceeds supply.

The key insight from The Great Mental Models, Volume 3 is that most healthy systems maintain dynamic equilibrium — not a frozen point, but a constant dance within a viable range. Pushing too hard for a fixed outcome can break the very balancing loops that were sustaining the system. Rent control is the classic case: the policy is intended to reduce the first-order price, but by disrupting the feedback between rents and new construction supply, it eventually shrinks the stock of available housing, producing the opposite of what was intended.

Feedback loops and the map-territory gap

One of the more consequential applications of feedback-loop thinking appears in The Great Mental Models, Volume 1's topic on maps and territories: feedback is the only mechanism that can correct a model. When you mistake the map for the territory — treating a dashboard, a report, or a theory as if it were the thing being described — you implicitly disable the feedback loop. You stop seeking disconfirming information. The map drifts away from reality while remaining internally consistent, so nothing alerts you to the divergence.

Jeff Bezos's practice of occasionally dialing Amazon's customer-service line himself, even when dashboards reported short wait times, was an act of forcing direct feedback past an organizational filter. The dashboard said under sixty seconds; his direct experience said over ten minutes. The feedback loop that would have corrected the model had been blocked by a layer of summarization. This pattern — a measurement system that systematically under-reports the thing it is supposed to measure, combined with a culture that trusts the measurement — produces the most insidious kind of map-territory drift.

The lesson is structural: good systems deliberately maintain multiple, redundant feedback channels, including at least some that are direct, unmediated, and hard to game.

Feedback loops and second-order thinking

Second-order thinking — the discipline of asking 'and then what?' — is essentially a practice of tracing feedback loops before they play out. When a policy, decision, or action is analyzed only to its first-order effect, the reinforcing or balancing loops it creates are left invisible. Those loops then play out in the world, producing effects the original decision-maker attributes to bad luck or external factors.

Rent control again: the first-order effect (lower rents for existing tenants) is real and arrives quickly. The second-order effects — reduced construction, reduced maintenance, conversion of units to condos — are the balancing loops fighting back against the imposed constraint. The third-order effect is a frozen, decaying housing stock that prices out new residents for a generation. The entire causal chain is a feedback loop analyzed by second-order thinking.

The asymmetry here matters: reinforcing loops make early wins appear to validate a decision even as second-order consequences accumulate silently. A bonus structure that rewards a single metric delivers real gains at first; the loop only visibly breaks when the unmeasured work it has been crowding out finally collapses.

Feedback loops and the circle of competence

A subtler application from The Great Mental Models, Volume 1 connects feedback loops to the circle of competence. Real circles of competence are built through feedback — through being wrong publicly, updating your model, and cycling through enough iterations to internalize failure modes. Imagined circles are tested only against your own confidence, which is closed-loop: your beliefs generate your judgment of your beliefs, with no outside signal arriving to correct the error.

This matters because feedback latency differs dramatically across domains. In chess or poker, feedback arrives quickly and unambiguously. In investing, management, or policy, the feedback loop between decision and outcome can span years, and many confounding variables cloud the signal. In those domains, building a genuine circle requires actively engineering feedback: keeping written records of predictions and revisiting them, seeking environments where wrong answers surface rather than get buried, and resisting the social pressure to attribute every good outcome to skill and every bad one to circumstances.

Author's argument (Parrish, Vol. 1): A track record tested against the world is the only honest way to know whether you have a real circle of competence or a comfortable-feeling map of unknown accuracy. The loop between judgment and outcome is what calibrates confidence to reality.

Practical levers: speed, gain, and rewiring

Once you have identified the loops in a situation, there are three basic interventions:

Feedback in human and social systems

Adam Smith's argument in The Theory of Moral Sentiments — that civility persists because social approval and disapproval shape future behavior — is a description of a slow-moving reinforcing loop. Each cooperative act generates more approval, which increases the probability of more cooperative acts, which builds social trust, which makes cooperation cheaper, which produces more of it. The entire fabric of stable civic life is an emergent property of billions of interlocking balancing and reinforcing loops operating across individuals and institutions.

This is also why interventions in social systems are so difficult to predict: the loops are long, overlapping, and partially invisible. Changing one variable changes the gain of several loops simultaneously, and the emergent effects — the system-level behaviors that arise from the interactions rather than from any single loop — cannot be read off from the components. The appropriate response, as Donella Meadows argued and as the Volume 3 topic echoes, is to move thoughtfully, watch what emerges, and be willing to revise.

Mental model

Where it goes next