Jon Moshier
Jon Moshier / Notes / Systems Thinking draft
Note · From the Notebook

Systems Thinking

How feedback structure generates the behavior of complex systems, why intuitive interventions backfire, and where leverage actually lives.

draft Started June 6, 2026 complexityscienceplanning

Systems Thinking

Systems thinking treats behavior as a product of structure rather than of events or actors. The recurring lesson across its history is uncomfortable: the intervention that looks obviously correct is frequently the one that makes the problem worse, because the human mind models causes as linear and immediate while real systems run on feedback and delay.

The founding discovery

In 1956, Jay Wright Forrester — an electrical engineer who had invented magnetic-core memory at MIT — moved to the Sloan School of Management and met managers from General Electric’s Kentucky appliance plants. The plants cycled violently: full capacity with overtime, then layoffs two or three years later. Executives blamed the external business cycle. Forrester was skeptical.

He simulated the factory’s ordering rules and time delays by hand, on a single notebook page. The result: management’s own method of projecting future demand, combined with delays in the manufacturing pipeline, was sufficient to generate the instability. No external shock required. Small shifts in demand produced disproportionate swings in production and employment. That notebook page was the first system dynamics simulation, and the founding act of the discipline.

The phenomenon became the Bullwhip Effect: small variations in end-customer demand amplify as they propagate upstream, because each tier over-orders to cover a pipeline of in-transit orders it cannot see. John Sterman (1989) reproduced it experimentally with MIT Sloan students playing a four-tier supply chain under constant end demand. Players still generated large inventory oscillations. He attributed it to “misperceptions of feedback” — decision-makers ignore the orders already on the way and keep ordering.

Stocks, flows, and the cognitive failure

The vocabulary is minimal.

A stock is anything that accumulates: CO₂ in the atmosphere, water in a bathtub, confidence in a person. A stock is the memory of a system; it carries the history of every past flow. Stocks change slowly even when flows change fast, which is the origin of systemic inertia.

A flow is a rate into or out of a stock: births and deaths for a population, income and spending for a bank balance.

Feedback loops close the chain — a change in a stock alters flows, which alter the stock. Balancing loops are goal-seeking; they resist change and create stability (a thermostat, predator-prey regulation). Reinforcing loops are self-amplifying; they produce exponential growth or collapse (compound interest, viral spread).

Delays are where intuition breaks. A gap in time between action and effect makes a system overshoot its corrections and oscillate, because the operator is acting on stale information. Turn up the shower, feel nothing, turn it up more, get scalded, overcorrect cold. Interest-rate changes take 12 to 18 months to move inflation. Forrester’s deeper point: delay alone, with no external variation, is enough to generate the bullwhip.

The failure is measurable. Booth Sweeney and Sterman (2000) gave MIT Sloan graduate students graphs of a bathtub’s inflow and outflow and asked them to sketch the water level. Fewer than half got it right. The dominant error was the “correlation heuristic” — assuming the stock tracks the shape of the net flow rather than its integral. People who can solve the underlying calculus on paper still fail the intuition. This is why verbal systems language is no substitute for an actual model.

Leverage points

Donella Meadows — lead author of The Limits to Growth, MacArthur fellow, Dartmouth — published “Leverage Points: Places to Intervene in a System” in Whole Earth in 1997. It ranks 12 intervention types from weakest (12) to most powerful (1):

12–10 (weakest) — parameters and physical structure: Constants, taxes, subsidies, standards, buffer sizes, infrastructure. Nearly all policy debate lives here. Adjusting a parameter rarely changes behavior, because the feedback structure underneath is untouched. You cannot retroactively change a population’s age structure: “five-year-olds become six-year-olds predictably and unstoppably.”

9–7 — feedbacks: Lengths of delays, strength of balancing loops, gain on reinforcing loops. Meadows’ example: show homeowners a real-time display of electricity cost and usage drops — a missing feedback loop made visible. Often it is more powerful to take your foot off the accelerator (slow a reinforcing loop) than to strengthen the brake (a balancing loop).

6–4 — design: Who has access to which information, the rules of the system, the system’s capacity to reorganize itself. Rules (lever 5) shape the entire incentive landscape; biological evolution — the power to generate new structure — is lever 4, and diversity is its raw material.

3–1 (most powerful) — intent: The goal of the system (3), the paradigm from which goals and structure arise (2), and the capacity to hold any paradigm lightly (1). A corporation organized around quarterly profit cannot be tuned into one that protects an ecosystem by adjusting its loops; the goal cascades down through every structure below it.

Her warning sits at the center of the field: “The most frustrating aspect of high leverage points is that they are often unintuitive. Pushing harder in the direction that seems obvious often makes things worse.” High leverage points move systems hard, but not reliably in the direction the intervenor intended.

Abson et al. (2017) tested this against real sustainability policy and found interventions cluster overwhelmingly at the shallow parameters tier. The deep tier — goals and paradigms — is rarely targeted: harder to fund, harder to measure, politically contested. The framework names where leverage lives without solving the political economy of why no one goes there.

Limits to Growth: the field’s most contested model

In 1970 Forrester sketched a model of the global economy and ecosystem on a flight home from a Club of Rome meeting. A Dartmouth-bound MIT team led by Dennis Meadows, with Donella Meadows, turned it into World3: five interlinked sectors — population, food, industrial output, nonrenewable resources, pollution — and the 1972 book The Limits to Growth.

The central finding is a mechanism, not a forecast. Exponential growth inside a finite system, governed by reinforcing loops and delayed corrective feedback, does not glide to a soft landing. It overshoots, because by the time scarcity is unmistakable the system is already past sustainable levels and the brakes act too late. The “standard run” scenario showed growth through the early 21st century followed by overshoot and decline around mid-century.

The backlash was severe and frequently dishonest. Economist Wilfred Beckerman called the book “a brazen, impudent piece of nonsense”; William Nordhaus’s “Lethal Models” (1993) cemented the mainstream verdict that its predictions had failed. The problem, documented carefully since, is that the book modeled scenarios rather than issuing the dated resource-depletion predictions critics attacked. Most of the famous refutations targeted claims the text never made.

The empirical twist: Graham Turner (2008), a physicist at CSIRO, compared the standard run against 30 years of observed data and found they tracked closely, “which indicates the early stages of collapse occurring in about 2015.” His 2014 update and a separate University of Melbourne analysis confirmed it; a 2024 recalibration with 50 years of data found the business-as-usual trajectory broadly intact. The model that was declared dead in 1993 turned out to match reality better than its critics did. Whether it continues to is the live question.

System traps

Meadows catalogued recurring structures that produce predictable pathologies:

Policy resistance (fixes that fail): A short-term fix suppresses a symptom and spawns side effects that recreate it. Drug enforcement raises prices, which raises profit margins, which expands supply — the fix funds the problem.

Shifting the burden: The symptomatic solution feels effective and gets repeated while the fundamental one atrophies from disuse. Staffing around a broken process instead of fixing it. Addiction generalizes the pattern.

Escalation: Each actor responds to the other by raising its own action. Arms races, price wars. Every move is locally rational; the aggregate is destructive.

Success to the successful: A slight lead yields more resources, which yields a larger lead. Network effects, prestige cycles, platform monopolies. “To him who has, more will be given.”

[private link]: Shared resource, individually rational use, collective depletion, with the cost delayed and spread across everyone. Garrett Hardin named it in Science (1968) and treated collapse as near-inevitable. Elinor Ostrom’s Nobel-winning Governing the Commons (1990) refuted the inevitability empirically — fisheries, Spanish irrigation, Swiss and Japanese mountain villages — showing commons survive when communities build their own governance. Tragedy is the outcome when governance is absent, not a law of nature.

Where it fails

Every model requires a human to decide what is inside the system and what is outside — and that boundary is a value judgment, not a technical one. Critical Systems Heuristics, developed by Werner Ulrich from 1987, supplies 12 boundary questions to force those judgments into the open. Forrester’s Urban Dynamics (1969) is the cautionary case: modeling a city without race, redlining, or migration produced the “counterintuitive” conclusion that building low-income housing deepens urban poverty. The finding was less a discovery than a restatement of the assumptions, and critics in IEEE and elsewhere said so at the time.

Two further failure modes. A 2023 study found that running network-centrality measures over causal loop diagrams systematically misidentifies leverage points — the visually central variable is often not the dynamically powerful one, so the diagram breeds false confidence. And the language of feedback can quietly depoliticize: if outcomes are “systemic,” the question of who built the structure and who profits can vanish. Sterman’s own “All Models Are Wrong” (2002) makes the internal version of the critique: models are instruments for learning, not descriptions of reality, and practitioners routinely communicate outputs with more confidence than the model earns.

Try it

The Beer Game (2–3 hours, any 4 people). Run the supply-chain simulation: four roles, a simple demand signal, no communication between tiers. A free web version exists. Watch whether your group can suppress the oscillation even after being told the trap exists. Most cannot — which is the result, not a failure of play.

Bathtub exercise (30 minutes, paper). Draw an inflow curve and an outflow curve for a stock that fills then drains. Hand the two flow graphs to someone and ask them to sketch the stock level before you explain anything. Look for the correlation heuristic — a stock drawn to mimic the flow’s shape instead of its accumulation. Fewer than half of MIT Sloan graduates integrate correctly on the first try.

See also

Sources

← All notes Read recent essays →