Here is a true thing that should not be possible. You have a road network where everyone drives selfishly — each person takes whatever route gets them home fastest, given what everyone else is doing. The system has settled into a stable pattern. Then the city opens a brand-new road: free to use, with no toll, no flaw, no construction defect, faster than the roads around it. A pure gift of capacity.
And every single driver's commute gets longer.
Not because of the construction. Not because of an accident. The new road works exactly as designed — and precisely because it works, everyone rushes onto it, and the whole network jams worse than before. Worse still: no driver can fix it alone. Each one, asked “would you be faster on a different route?”, honestly answers no. They are all individually doing the best they can, and they are all collectively worse off than if the road had never been built. This is Braess's paradox (Dietrich Braess, 1968), and the network below is the real thing, computed live. Build the road. Watch it happen.
Take the textbook setting (the one in every algorithmic-game-theory course; the round figures are Easley & Kleinberg's, not Braess's original parametrisation). 4000 drivers leave Start for End. Two roads are congestible — the more cars on them, the slower they go, costing x/100 minutes when x cars are present. Two are highways at a flat 45 minutes, immune to traffic.
With no shortcut, the network is symmetric, so the drivers split 2000 / 2000 between the two L-shaped routes. Each route then costs 2000/100 + 45 = 65 minutes, and nobody can do better by switching — a Wardrop equilibrium, the traffic version of a Nash equilibrium.
Now add a shortcut from the upper node to the lower one, so fast we'll call it free (0 minutes). It opens a new route: Start → upper → across the shortcut → lower → End, using both congestible roads and skipping both highways. When traffic is light that route is irresistible — so everyone takes it. But if all 4000 pile onto both congestible roads, each one now costs 4000/100 = 40 minutes, and the zig-zag route totals 40 + 0 + 40 = 80 minutes.
80 > 65. Everyone is slower. And it is a genuine trap: if a lone driver defects back to an old L-route, they pay 40 + 45 = 85 minutes — worse than the 80 they're suffering. No one can escape unilaterally. The only fix is collective, and selfish routing can't reach it. The 4000 figure is incidental; move the sliders and the paradox appears, peaks, and vanishes on its own schedule.
The shortcut is not always poison. Push the highway time slider up and you'll watch the paradox switch off — and then reverse. With congestible slope s = 1/100, demand N, and highways costing c, the selfish equilibrium with the shortcut open is exactly:
| highway cost c | what the shortcut does | per-driver time |
|---|---|---|
| c ≤ N/200 | ignored — nobody uses it | N/200 + c |
| N/200 < c < N/100 | partly used (a mix) | 2c |
| c ≥ N/100 | everyone floods it | N/50 |
Compare that to the no-shortcut time, always N/200 + c. The road strictly hurts exactly when c < 3N/200 — cheap highways relative to the crowd. When the highways are slow enough (c > 3N/200) the very same shortcut helps. At N = 4000 the tipping point is c = 60 minutes: below it, building the road is a mistake; above it, a kindness. Every cell of that table is checked against an independent network solver across 216 parameter combinations.
Braess's paradox is the dramatic face of a quieter quantity: the price of anarchy — how much worse a selfish equilibrium is than the route assignment a benevolent planner would impose. Here is a subtlety the folklore usually gets wrong: in the 4000-driver network the planner does not simply ban the shortcut and split 2000/2000. The true optimum keeps 500 drivers on the zig-zag and routes 1750 down each L-route, for 64.6875 minutes a head — a hair better than 65. The selfish crowd's 80 is therefore 80 / 64.6875 = 256/207 ≈ 1.237 times the optimum.
That ratio can't run away. Roughgarden and Tardos proved (2002) that for any network with linear road-cost functions, the price of anarchy is at most 4/3 — selfish routing is never more than 33% worse than optimal, no matter how the network is wired. The bound is tight: the normalised unit-demand version of this very network (1.5 → 2.0) hits exactly 4/3. So this paradox is, in a precise sense, as bad as linear traffic ever gets.
The drivers were a costume. The skeleton is: independent agents each minimising their own cost, coupled through shared congestible resources — and that skeleton wears many bodies. In 1991, Joel Cohen and Paul Horowitz built a contraption of springs and strings holding up a weight, rigged so that cutting a string — removing a connection — makes the weight rise, and re-tying it makes the weight sink. Same paradox, run in reverse, in steel and cord (Nature 352, 699, 1991). They give the electrical twin too: a circuit where adding a wire raises the total voltage drop. In 2012 a group led by M. G. Pala saw it in electrons — adding a third channel to a tiny semiconductor network cut the current that flowed through (Phys. Rev. Lett. 108, 076802). The paradox is a property of networks of self-interested flow, not of asphalt.
This is where honesty earns its keep, because the famous real-world stories are shakier than they're usually told. The cleanest demonstrations of Braess's paradox are the mathematics above, the controlled physical analogues, and computer models of real street maps. The on-the-ground anecdotes are suggestive but confounded — when you close a road, some trips simply don't happen, which lowers congestion for reasons that aren't the paradox at all. Here is what each claim actually rests on:
New York, 1990 cited — closing 42nd Street for Earth Day was expected to snarl traffic; it eased instead. Source: Gina Kolata, “What if They Closed 42d Street and Nobody Noticed?”, The New York Times, 25 Dec 1990. (Citation confirmed; the primary text is paywalled, so no sentence is quoted here verbatim.)
Seoul, 2003–2005 partly verified — the Cheonggye elevated expressway (~6 lanes, ~168,000 vehicles/day) was torn down for the Cheonggyecheon stream restoration; traffic did not collapse as feared. Demolition began July 2003; the stream reopened 1 Oct 2005 (City of Seoul). But the transport literature attributes much of the effect to reduced and re-timed demand, not pure rerouting — so calling it strict Braess is an interpretation, not a measurement.
Stuttgart, ~1969 anecdotal — the oft-repeated tale that a new downtown road only helped once it was closed again traces to a 1969 graph-theory textbook (W. Knödel), not a measured field study. No traffic data substantiates it. Repeated often; demonstrated never.
(A 2008 study by Youn, Gastner & Jeong did find Braess-prone roads in network models of Boston, New York, and London — but it models those three cities, not Seoul, and is a model, not a before/after road measurement.)
So: the paradox is real, exact, and provable — and the everyday version of “more capacity must mean less congestion” is simply false as a theorem. But the dramatic newspaper instances are best held as plausible illustrations, not clean proofs. The certainty lives in the network you can build above, where every number is recomputed and nothing is taken on faith.