> ## Content Index
> Fetch the complete content index at: https://machina.mmalc.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Traffic jams, and why you’re in one
- URL: https://machina.mmalc.com/traffic-jams/
- Published: 2026-08-31T04:53:54.000Z
- Updated: 2026-08-31T21:36:34.000Z
- Description: How they form without anyone causing them, why going faster buys less than it seems, and two things worth changing on the next journey.
- Author: mmalc Crawford
- Tags: traffic

## Motorway traffic does not behave the way it appears to

Some things about motorway driving seem too obvious to check. A driver who gets past more vehicles arrives sooner. A queue has a cause, and somebody at the front of it is responsible.

Both are reasonable inferences from ordinary experience. Both also fail in the situation that matters most — the approach to a queue — and they fail in ways that change what a driver should do. A third finding is less obvious but just as useful: adaptive cruise control does not necessarily make the traffic around it smoother.

Together, the findings suggest two specific changes for the next journey: lift off when a queue becomes visible, and, when using adaptive cruise control, select the longest following-distance setting. This article explains the science behind the counter-intuitive findings, and why the two suggested behaviours work.

## Ten miles an hour buys less at the top of the range

The familiar speedometer seems simple enough: a higher number means going faster and arriving sooner. But speed and time saved are related by a curve, not a straight line, and human intuition handles that conversion badly. What matters to a driver deciding how hard to push is the inverse of miles per hour: minutes per mile.

The consequence for a driver approaching congestion is that lifting off feels more costly than it is.

The Swedish psychologist Ola Svenson identified this problem in the 1970s. It became known as the time-saving bias: people systematically overestimate the time saved by increasing an already-high speed and underestimate the saving available at lower speeds. Experience does not reliably correct the error.

The arithmetic is straightforward. On a 100-mile journey, raising the average speed from 50 to 60 mph saves twenty minutes. Raising it from 70 to 80 mph saves less than eleven. The same increase, applied at the top of the range, buys roughly half as much.

Distance scales the saving proportionally, and most journeys are short. Over thirty miles, the difference between averaging 60 and 70 mph is a little over four minutes. Expressed as pace, 60 mph takes ten minutes per ten miles, 70 takes about eight and a half, and 80 takes seven and a half. A substantial increase in speed from 70 to 80 returns only about one minute every ten miles.

On the approach to congestion, the calculation becomes particularly stark. Covering the final half-mile at 70 mph rather than averaging 50 saves only about ten seconds. If the traffic ahead is still stopped, those seconds are simply spent waiting at the back of the queue. Reaching the back of the queue sooner is not the same as getting through it sooner.

## A queue does not need a culprit

A traffic jam invites a search for a culprit — a collision, roadworks, someone driving badly at the front of it. Often there is none to find. The queue clears with nothing to explain it, having been produced by everybody in it and by nobody in particular.

Something always begins the disturbance, but it need not be a mistake. This was illustrated vividly in a 2008 experiment in which a team led by Yuki Sugiyama put twenty-two vehicles on a circular track and asked their drivers to maintain a constant speed. There were no junctions, obstructions, or lane changes. Just ordinary drivers driving normally, with the small variations that normal driving entails. Within minutes, a traffic jam had formed — a stop-and-go wave travelling backwards around the ring.

On a normal motorway, a driver may ease off slightly to restore a safe following gap. That is an entirely defensible response. The problem develops as the vehicles behind respond to that change, because no car-following response is instantaneous. Even an attentive human driver must perceive a change, decide how to respond, and act; an adaptive cruise-control system must sense, compute, and actuate. In both cases, the response comes after the change ahead and does not reproduce it exactly.

In dense traffic, those small delays and corrections can make a disturbance grow as it travels upstream. Traffic engineers call this string instability. Drivers encounter it as a phantom jam: traffic stops, starts, and eventually clears without any accident or roadworks appearing.

## Arrive as the queue starts moving

A common response to congestion is to maintain speed until the brake lights immediately ahead demand action, then brake, stop, and accelerate when the queue moves. That passes the stop-and-go wave backwards, and may amplify it.

The alternative begins as soon as stopped or stop-start traffic becomes visible. Lift off, shed speed gradually, and allow a buffer to open. The aim is to reach the vehicle ahead after it has begun moving again, and especially to avoid coming to a complete stop.

In traffic research, this is known as jam-absorption driving. A driver with space available can continue at a steady low speed while the vehicles ahead alternate between moving faster and standing still. The absorbing vehicle does not overtake the queue or receive priority. It simply stops copying the oscillation, so the wave leaving it can be smaller than the wave that reached it.

![Two animated lines of cars approach the same slowdown. In the upper line, successive drivers brake later and the stop-and-go wave travels backwards. In the lower line, a highlighted driver lifts off early, creates a buffer, and helps the following traffic keep rolling.](https://storage.ghost.io/c/8d/68/8d6859a9-24e5-4568-91fa-cbd411bb8e53/content/images/2026/08/phantom-jam-intervention-with-slow-replay.gif)

The same disturbance, with two responses. Reacting only to the vehicle immediately ahead passes the stop-and-go wave backwards. Looking further ahead, lifting off early, and preserving a buffer can make the wave smaller.

If the vehicle ahead is already doing this, leave enough room to follow the same pattern. The absorbing driver cannot maintain an exact speed indefinitely; traffic ahead may compress and force it to slow further, or even to stop. A driver following too closely will then have to stop as well, reintroducing the wave. A driver who has kept a buffer may still be able to keep rolling. It costs no progress through the queue, and it helps the vehicles behind.

The effect matters because a stopped queue cannot move off simultaneously. Each driver waits to perceive movement ahead, react, and begin accelerating. The resulting start-up headway is roughly a second or two per vehicle, and it accumulates down the line. A queue of two hundred stopped cars therefore takes several minutes to restart from front to back, assembled from delays contributed by drivers who individually did nothing wrong.

A driver who keeps rolling contributes none of that start-up delay.

The buffer is not sacred territory. Another vehicle may enter it, especially from an adjacent lane. Let it. Braking to defend the space creates exactly the disturbance the gap was meant to absorb. Re-open it gradually instead.

## Adaptive cruise control is not yet on the driver's side

Adaptive cruise control seems well suited to smooth traffic. It holds a selected following distance without impatience or fatigue. But consistent gap control is not the same thing as string stability, and measurements of production systems do not support the assumption that one guarantees the other.

In more than 1,200 miles of car-following experiments, all seven 2018-model-year vehicles tested were string unstable: changes in speed became larger as they travelled backwards through a line of vehicles. In a platoon test using one of those models, a 6 mph disturbance at the front grew to 25 mph by the last vehicle.

Other experiments with commercial systems have found the same underlying problems. Their responses are not instantaneous, and the way they close a gap can amplify a change rather than absorb it. Most regulate their relationship with the vehicle immediately ahead; they do not respond as a human anticipatory driver can to a queue visible several vehicles or half a mile further on.

Following distance is the control the driver does have. Longer headways improve the conditions for string stability; shorter ones make amplification more likely. The precise threshold varies by controller, speed, and disturbance, and manufacturers do not publish which setting, if any, crosses it for a particular model.

The usable advice is therefore simple: choose the longest following-distance setting, every time.

That setting may invite other vehicles into the gap, and a longer headway uses more road space. Those are real trade-offs. Neither changes the direction of the advice: if the aim is to avoid amplifying disturbances, the longest available setting gives the system its best chance.

## A small fraction can change the stream

The same research offers a useful closing result. In closed-track experiments with more than twenty vehicles, putting a single vehicle under control designed to damp the wave reduced the oscillations across the group and cut collective fuel consumption by as much as roughly 40 per cent.

![Twenty-two cars travel clockwise around a circular track. Ordinary variations grow into a red stop-and-go wave moving backwards around the ring. The sequence then repeats with one blue smoothing vehicle, which opens a buffer before the wave arrives and causes the disturbance to shrink.](https://storage.ghost.io/c/8d/68/8d6859a9-24e5-4568-91fa-cbd411bb8e53/content/images/2026/08/circular-phantom-jam.gif)

Cars travel clockwise while the stop-and-go wave travels backwards. The first sequence illustrates the 2008 circular-track experiment; the second illustrates later experiments in which one smoothing vehicle reduced the disturbance across the stream.

The approach has also been tested in live traffic. In 2022, researchers released one hundred vehicles running experimental cruise-control software into the morning commute on I-24 near Nashville. It tested whether smoothing control can influence real traffic without waiting for an entirely automated fleet, in conditions a ring track cannot reproduce.

The significance is not that ordinary drivers can reproduce an experimental controller exactly. It is that a small proportion of vehicles behaving differently can change the stream around them. A few drivers lifting off early, preserving a buffer, and declining to pass the full wave backwards are not merely being considerate. They are changing the traffic they themselves will occupy.

## The considerate driver and the self-interested driver should do the same two things

The competitive account of motorway driving assumes that time gained by one driver is time taken from another, and that restraint is a cost paid to strangers.

The arithmetic and the traffic dynamics say otherwise. In the final approach to congestion, driving harder buys seconds, not meaningful progress. Reaching a queue sooner does not get anyone through it sooner. Once a disturbance begins, each delayed or excessive response can enlarge the wave, and the driver who helps enlarge it may shortly be sitting in it.

The considerate driver and the purely self-interested driver should therefore do the same two things: lift off when a queue becomes visible, and give adaptive cruise control the longest gap available. Neither requires an appeal to patience or virtue. They are simply the better responses to the way motorway traffic actually behaves.

## Evidence notes

1. **The time-saving bias.** Ola Svenson's early experiments established that people judge the time saved by speed increases as though the relationship were much closer to linear than it is. Eyal Peer and Eyal Gamliel proposed replacing the speedometer with a pace display as a corrective. A simulator study published later the same year found that the bias survives active driving, and a subsequent experiment found that displaying inverse speed brought drivers closer to a target time saving. A separate study confirms the mirror case, in which drivers misjudge the time lost by slowing down: [Svenson (1970)](https://doi.org/10.1037/h0029934?ref=machina.mmalc.com), [Peer and Gamliel (2013)](https://doi.org/10.1017/S1930297500005040?ref=machina.mmalc.com), [Eriksson, Svenson, and Eriksson (2013)](https://doi.org/10.1017/S1930297500005337?ref=machina.mmalc.com), [Eriksson et al. (2015)](https://doi.org/10.1080/00140139.2015.1051592?ref=machina.mmalc.com), and [Svenson and Treurniet (2017)](https://doi.org/10.1016/j.trf.2017.09.007?ref=machina.mmalc.com).
2. **Phantom jams without a bottleneck.** Sugiyama and colleagues asked twenty-two drivers on a circular track to maintain the same moderate speed. The initially even flow broke into a backward-moving stop-and-go wave without an accident, merge, lane change, or obstruction: [*Traffic jams without bottlenecks — experimental evidence for the physical mechanism of the formation of a jam* (2008)](https://doi.org/10.1088/1367-2630/10/3/033001?ref=machina.mmalc.com).
3. **Jam-absorption driving.** The theoretical work describes how a single vehicle can use a "slow-in" phase to create headway upstream of a jam, cutting the supply of vehicles reaching it, and identifies conditions intended to prevent the manoeuvre from creating a secondary jam. Later modelling added realistic acceleration and car-following behaviour: [Nishi et al. (2013)](https://doi.org/10.1016/j.trb.2013.02.003?ref=machina.mmalc.com) and [Taniguchi et al. (2015)](https://doi.org/10.1016/j.physa.2015.03.036?ref=machina.mmalc.com).
4. **Queue discharge.** Empirical studies of moving jams place the mean delay before successive vehicles accelerate at roughly 1.3 to 2.1 seconds. This is the microscopic delay behind the start-up arithmetic in the article: [Rehborn, Klenov, and Palmer (2011)](https://doi.org/10.1016/j.physa.2011.07.004?ref=machina.mmalc.com).
5. **Commercial adaptive cruise control.** Seven production vehicles from two manufacturers were tested over more than 1,200 miles with ACC engaged. Every fitted system was string unstable; in the validating platoon experiment, a 6 mph disturbance grew to 25 mph by the final vehicle: [Gunter et al. (2021)](https://doi.org/10.1109/TITS.2020.3000682?ref=machina.mmalc.com). The wider [OpenACC dataset](https://arxiv.org/abs/2004.06342?ref=machina.mmalc.com) assembled experiments involving commercial systems from multiple manufacturers.
6. **Following-distance settings.** Closed-course experiments with commercial automated car-following tested different selectable headways and found that stability depends materially on the setting, with shorter headways trading stability for capacity: [Shi and Li (2021)](https://doi.org/10.1016/j.trc.2021.103134?ref=machina.mmalc.com). The longest setting is not a guarantee: in the Gunter experiments, simulations of one vehicle model at its maximum following setting remained string unstable, though the disturbance grew far more slowly than at the minimum setting.
7. **Looking beyond the first vehicle.** Modelling based on commercial ACC behaviour found that monitoring two vehicles ahead can provide string stability across a wide range of conditions without vehicle-to-vehicle communication: [*Multianticipation for string stable Adaptive Cruise Control and increased motorway capacity without vehicle-to-vehicle communication* (2022)](https://doi.org/10.1016/j.trc.2022.103687?ref=machina.mmalc.com).
8. **One controlled vehicle on a closed track.** Experiments used more than twenty vehicles on a circular track and changed the control of one. That one vehicle could damp the stop-and-go wave across the group, and the authors conclude that fewer than 5 per cent of vehicles under control would suffice to influence flow; fuel consumption fell by 42.5, 22.1, and 28.1 per cent across the three experiments: [Stern et al. (2018)](https://doi.org/10.1016/j.trc.2018.02.005?ref=machina.mmalc.com).
9. **One hundred vehicles in live traffic.** The CIRCLES field experiment deployed one hundred control-equipped vehicles on I-24 in ordinary multilane traffic, with the surrounding flow measured by roadside infrastructure. It tested whether sparse vehicle control could smooth real traffic outside the simplified conditions of a ring track: [Lee et al. (2025)](https://doi.org/10.1109/MCS.2024.3498552?ref=machina.mmalc.com).

## Short visual explanations

**Why speed saves less time than it appears to.** Rory Sutherland makes the arithmetic intuitive using a paceometer — a dashboard readout proposed by Eyal Peer and Eyal Gamliel in 2013 that shows the minutes required to cover a given distance instead of miles per hour. The uneven scale makes immediately visible why adding speed at the upper end buys so little time. [Watch the presentation extract.](https://youtube.com/shorts/fBfchhyY5Q0?ref=machina.mmalc.com)

**How a phantom jam forms.** Hannah Fry gives a clear visual explanation of how a small disturbance grows into a stop-and-go wave travelling backwards through traffic. [Watch the explanation.](https://youtube.com/shorts/-1Bv71mKocI?ref=machina.mmalc.com) 

Her advice to a driver already caught in one — sit back and let it wash past — is right for the question she is answering. A wave that has arrived cannot be escaped, and nothing that driver does will get them out of it any sooner. But that question is about the driver's own predicament, and a driver's predicament is not the whole of what they affect. Slowing early and holding a buffer will not clear the queue ahead; it changes the wave handed to everyone behind.

Her two remedies — driverless cars with faster reactions, or universal agreement to stop tailgating — share the assumption that any fix would have to be universal, and that is where the disagreement lies. Reaction speed is not the variable: string stability depends on how a system closes a gap and what headway it holds, which is why all seven production adaptive cruise control systems studied by Gunter and colleagues were found to be string unstable. And the proportion of vehicles required is small rather than universal.

---

*Researched and drafted in collaboration with Claude (Anthropic).*