communicoupling · concept 10 of 26

A city sorts itself
that no one wanted.

How does a city end up sharply segregated when almost no one in it wants to be? Schelling's model shows the gap between what individuals prefer and what they collectively produce: a mild taste for not being too outnumbered, iterated across a whole population, tips into near-total separation no single person intended.

Ask people directly and few want segregation; most say they would happily live in a mixed neighbourhood, and mean it. Yet mixed neighbourhoods are rare, and the ones that exist feel fragile. The reflex is to reach for prejudice as the explanation — and prejudice is real, and does its damage. But Schelling showed you don't even need it. A slight preference, held by everyone, for not being stranded in too small a local minority is already enough to pull a whole map apart.

The macro-pattern betrays the micro-motive. The finished map looks like the work of people who hated each other, but it can be assembled entirely by people who would each have been content in a mix — they just each made one small, reasonable move, and the moves added up to something none of them was aiming at.

colour A colour B prefers-mixing empty
Currently unhappy
Segregation index
The gap that betrays the motive
what each agent would accept
↓ yet they live in ↓
what they end up living in
Reseed to a random mix, then run. Watch the lived share overshoot the accepted one by far more than anyone asked for.
0 sweeps · not yet run
Tolerance the minimum same-colour share of neighbours I'll accept before I want to move
33%
0% · anything goes 50% 85% · near-purist
Vacancyshare of cells left empty (room to move)
10%
Colour ratioshare of agents that are colour A
50%
Prefers-mixingshare who move away from too much sameness
0%
Speedsweeps per animation frame while running
Watch what happens
Every number here is computed live from a real agent-based simulation — an agent is unhappy when the same-colour share of its 8 neighbours drops below its tolerance, and unhappy agents relocate. Nothing is scripted.

Micromotives and macrobehaviour

The rule each agent follows is not the pattern they make together.

Each agent checks its eight neighbours, counts how many share its colour, and asks one modest question: am I in too small a minority right here? If the same-colour share falls below its tolerance, it is mildly uncomfortable and moves somewhere it isn't. That is the entire micromotive. At 33% tolerance it is a person perfectly happy to be a one-third minority — outnumbered two to one — and only twitchy below that. By any ordinary standard, that person is tolerant.

Now iterate that same reasonable rule across a couple of thousand agents. When one agent leaves a too-lonely spot, it slightly raises the minority-share of the neighbours it left behind, nudging them to move; and it slightly raises the same-colour comfort of wherever it lands, anchoring others there. The moves are coupled. Each relocation is a small, local, defensible choice, but the choices feed each other, and the feedback runs only one way — toward clumping. The segregation index in the readout is the average same-colour share every agent actually experiences. It starts near the 50% of a random mix and climbs, unbidden, to 75–85%. Nobody dialled that number. It is the sum of a rule that never mentioned it.

What to try

Break it yourself.

01

Easygoing still segregates

Drop tolerance to 30% — everyone content as a one-third minority — hit Run, and watch the grid still curdle into stark blocks. Compare the accepted share (30%) with the lived share (often 80%+).

02

Find the tipping tolerance

Nudge tolerance down toward 20–25% and reseed a few times. Somewhere low the map stops tipping and mixing survives the run. Above that knife-edge, separation is nearly certain. The threshold is sharp and lower than intuition expects.

03

Add integration-preferrers

Raise prefers-mixing — a minority who move away from too much sameness. A few can hold pockets of mix open; but push tolerance up and watch how quickly even they are overwhelmed. Integration is achievable, and delicate.

The aggregate is not the intention

You cannot read a population's wishes off its pattern.

The lesson cuts both ways. A segregated map is not proof of hatred: the very same instrument produces it from agents who would each have accepted a heavy minority around them. And an integrated map is not proof of virtue: you can hold mixing with anchors, vacancy, and a low enough threshold, without anyone being a saint. The outcome and the motive have come loose from each other. What you can see — the pattern — under-determines what you want to know — the wishes that made it.

That gap is where a specific moral hazard lives. It is tempting to run the inference backwards: the neighbourhood is sorted, therefore the people wanted it sorted, therefore we may judge them (or excuse ourselves). Schelling's grid is a standing refutation of that shortcut. The aggregate is not the sum of intentions; it is what a web of coupled, individually-innocent choices settles into. Reading motive off outcome is a category error — and, worse, a convenient one, because it lets the pattern testify about people who never agreed to be its witnesses.

Design against the tip

What actually holds mixing.

Exhortation barely moves the micromotive: asking everyone to be a little more tolerant while leaving the mechanism intact just delays the tip. Thresholds: keeping the effective tolerance below the knife-edge, structurally, changes everything, because the system is not linear near it. Anchors: a modest share of prefers-mixing agents, or fixed institutions that don't relocate, can pin pockets of integration open. Vacancy and friction: when moving is costly or slow, the feedback loop loses the fuel it runs on. Mixing is not held by good intentions; it is held by the shape of the choice architecture people move within.

Real segregation is not only, or even mainly, Schelling dynamics. It is also produced by income, by lending and zoning, by discrimination that is deliberate and enforced, by a history that loaded the dice before anyone in the model made a single move. The grid shows that mild preference is sufficient to segregate; it never claims preference is all that's at work. Design that leans only on the Schelling lever — nudging thresholds while ignoring the money and the policy — will find the map stubbornly refusing to integrate, because the real machine has parts this one omits.

The mapping

Mechanism ↔ social life.

In the modelIn the world
an agent's toleranceThe mild preference each person actually holds — not to be stranded as too small a local minority.
a moveOne person's local, reasonable choice to live somewhere a little more comfortable.
the segregation indexThe collective pattern that emerges — the average sameness everyone ends up surrounded by.
the gap (accepted vs lived)How far the outcome overshoots the wish — the distance between what people would settle for and what they get.
the tipping toleranceThe knife-edge threshold where mixing either survives or collapses; small moves across it flip the whole map.
prefers-mixing agentsWhat it takes — and doesn't — to hold a mixed order open: anchors that resist the pull toward sameness.

Where it tears

Three honest failures.

Sufficiency is not necessity.

The model proves mild preference is enough to segregate — not that it is what actually did. Real separation is also produced by income, discrimination, lending, and policy that the pure grid omits entirely. Schelling explains that segregation can arise innocently; he must never be used to argue that any particular segregation therefore did, or to explain the coercion away.

"It's just Schelling dynamics" launders injustice.

The most dangerous use of this beautiful result is as an alibi. Reframing a segregation with real victims and real perpetrators as innocent arithmetic — emergent, no one's fault, just micromotives — converts an injustice into a fact of nature. The model earns its keep by exposing how outcomes detach from motives; it forfeits it the moment that detachment is used to retire the question of responsibility.

Two colours flatten a loaded world.

Real difference is not two interchangeable tokens on a blank grid. It carries histories that pre-sorted the map, resources that make some moves possible and others impossible, and meanings that make "minority" mean wildly different things depending on which side you're on. The grid starts everyone equal and mobile; the world never did. The dice were loaded before the first move.