Skip to content

Without the hype

How Ideas Spread: A Murmuration Model

Ideas
Science & tech
A toy model of 400 people who fly like a flock of starlings and pass ideas on. Start an idea and watch it spread and fade, then change the emotion, evidence, social media, influencers and how readily people go with the flow to see what the model’s rules do. It uses no data.
Published

19 November 2026

For years I have thought that ideas move through a society the way a turn moves through a murmuration of starlings. No bird leads: each copies the few birds nearest it, and a turn that starts with a few birds at one edge can sweep across the whole flock. This page is a working model of that thought. It is a crowd of 400 people who fly like starlings and pass ideas to the people they hear from: their nearest neighbours, their contacts online and the influencers they follow. Start an idea, change how the people think and how they are connected, and watch how far the idea spreads, how fast, and how long people hold it before they lose interest. The model uses no data. Its rules come from published models of flocks, rumours and opinion, set out under How the model works, so its numbers show what those rules do, not what real people do.

An idea spreading through a flock of 400 people

Each triangle in the sky below is a person, flying the way a starling does. Its colour shows what they think of the current idea, the four figures under the sky count them, and the two charts below the settings follow the idea round by round. Pick an experiment to start again from my settings for it, or move the settings yourself while the idea runs.

Experiments

Share of the crowd, by round
Who has heard, believes or opposes the idea

A round is one chance for every person who is passing the idea on to reach each of their contacts. Point at or tap the chart to read the shares at any round.

Share of the crowd, from side A to side B
Where people stand, before this idea and now

Each person stands somewhere between side A (left) and side B (right); the further from the centre, the more attached they are. An idea that suits one side moves those who take it up towards that side and those who argue against it the other way.

How each experiment turns out over many runs

One run of an idea is partly luck: who hears it first, who happens to be nearby, who loses interest. The table runs each experiment 24 times, on new crowds, and gives the averages. An idea “caught on” when at least 5% of the crowd believed it at once; the rounds to the peak and until it faded are averages of the runs where it caught on. Polarisation is how far people stand from the centre, on average, on a scale of 0 to 100, and the last column is its largest rise during the idea. Click an experiment’s name to watch it in the sky above.

The runs start when you scroll here.

Results for four of the experiments:

  • Outrage reaches its peak in about 13 rounds, three times as fast as the idea on the starting settings, but about as many people argue against it at some point (39%) as believe it at some point (43%), and polarisation rises by about 9 points. In the model, each believer moves a step towards side B and each opponent a step towards side A, so an emotional idea that suits one side moves both sides away from the centre.
  • Solid evidence, little emotion peaks later, after about 36 rounds, but nine in ten people come to believe it, and it lasts more than twice as long as the idea on the starting settings before it fades.
  • No social media gets the idea to about half the crowd, with the peak after about 57 rounds. With most contacts online (and six influencers), the peak comes after about 34 rounds, and the idea reaches about the same share of the crowd as on the starting settings (63% and 65%); it caught on in 19 of 24 runs, against 21 of 24.
  • Started by an influencer, with 30% of contacts online and eight influencers, an idea peaks in about 16 rounds and reaches more than four in five people. On the starting settings (20% online, four influencers), an idea started by an ordinary person takes about 40 rounds and reaches about two thirds.

How the share of contacts online changes the spread

The comparison below changes the share of each person’s contacts who are online, from none (everyone hears only from the people nearest them) to all (everyone hears only from people scattered across the crowd). It does this for two kinds of crowd: one with going with the flow set to 85%, where most people take an idea up once one contact has, and one with it set to 35%, where most wait until two or three have. Every other setting is as it is above.

The runs start when you scroll here.

Share of the crowd, average of 24 runs
Ever believed it, by share of contacts online
Average of the runs where it caught on
Rounds to the peak, by share of contacts online

Each point is the average of 24 runs with that share of contacts online, each starting from a random ordinary person. Point at or tap a chart to read its values.

In the crowd that goes with the flow (85%), moving contacts online halves the time to the peak, from about 30 rounds with nobody online to about 15 with 40% or more online. The share who ever believe it stays at about 90% whatever the share online: in this crowd the wave from neighbour to neighbour reaches nearly everyone in the end, and online contacts get it there sooner.

The crowd that waits for several contacts (going with the flow at 35%) behaves differently, and its runs vary far more: in some the idea goes nowhere, in others it reaches most of the crowd. With nobody online the idea reached about a third of the crowd on average, and with 40% or more of contacts online, between about 43% and 53%. Online contacts didn’t make it faster: when it caught on, the peak came after about 30 to 40 rounds whatever the share online.

This crowd doesn’t reproduce two published findings. In an online experiment, a new health behaviour spread further and faster through clustered networks than through random ones (Centola, 2010), and in models of ideas that need several sources, a few long links help but too many stop the spread (Centola, Eguíluz and Macy, 2007). Here, more online contacts never hurt. One reason may be that people here keep flying and meeting new neighbours, so an idea that stalls in one place can still find a second contact for someone later.

Polarisation under a run of emotional ideas

In the model, each person stands somewhere on a line from side A to side B. When someone takes up an idea that suits one side, they move a step towards that side; when they argue against it, they move a step away. The step is larger for a more emotional idea, and between ideas everyone drifts slowly back towards where they started.

Tick “Keep the news coming” with the Outrage experiment and watch the chart of where people stand: the two humps, one for each side, move outwards. In one run of 12 ideas on those settings, from alternate sides, polarisation rose from 61 to 78; another run will give other numbers. The same 12 ideas made calm and well evidenced (emotional charge 15%, evidence 90%) left it at 63. Each emotional idea wins over part of the side it suits and turns much of the other side against it, so both sides move outwards, and the next idea finds people further apart. Most people on both sides took up the calm ideas, so the steps towards each side cancelled out.

I wrote about how groups shape their members’ emotional reactions, in activism, religion and politics, in Emotional Priming and Collective Conditioning. This model turns some of those ideas into rules that a computer can run.

What the model leaves out

  • One idea at a time. Real people meet many ideas at once, and ideas compete for a limited amount of attention.
  • One line of opinion. Worldview is a single line from side A to side B; real views have many parts that need not line up.
  • Evidence as one number. Here everyone agrees on how good the evidence is. In real arguments, people disagree about the evidence itself.
  • Fixed feeds. The model picks each person’s online contacts once, by chance or from their own side. Real feeds change from day to day, and platforms rank posts in ways the model leaves out.
  • 400 people. A real society has millions, with many separate outbreaks of the same idea, and a few accounts followed by millions.
  • Rounds, not days. A round has no set length in real time. The model can compare one setting with another, not forecast how long a real idea lasts.
  • Flying as a picture of mixing. Moving through the flock stands for meeting new people. Real people change who they meet over months and years, not seconds.

How the model works

Most numbers below are choices I made for the model, set so that ideas sometimes catch on and sometimes don’t. Two rules borrow from measurements: the seven neighbours each bird steers by, and the direction, not the size, of the effect of emotion.

Flying

Each person steers by their seven nearest neighbours; starlings interact with six or seven (Ballerini and colleagues, 2008), using the three rules of Craig Reynolds’s “boids” (Reynolds, 1987): keep a little space from the nearest birds, fly in the same direction as the neighbours, and steer towards their middle. Birds also turn at random a little, independent thinkers more than others. Four more rules come from the people and the idea:

  • Everyone steers a little towards the middle of the sky and turns back near its edge, as starlings circle over a roost.
  • People steer towards neighbours on their own side, in proportion to how attached both are, so a strongly attached crowd flies in patches of one side or the other.
  • People who hold the idea steer towards others who hold it, and people on opposite sides of it keep apart, more strongly when they are attached to their side and the idea is emotional.
  • People passing the idea on lean towards the side it suits (to the right, towards side B, for an idea that suits neither), and those arguing against it lean the other way, like the few informed animals in Iain Couzin’s model of how a small part of a group can lead it (Couzin and colleagues, 2005).

The last two rules are how a divisive idea splits the flock into separate groups of believers and opponents. Each crowd flies for 100 rounds before its first idea starts, so the flock has formed.

Contacts

Each person hears from a set number of contacts (7 to start with). Some are the people nearest them in the flock, so they change as the flock moves. The rest are online: people anywhere in the crowd, chosen once, at random or, as feeds match like with like, from the person’s own side. About a quarter of the crowd follows each influencer, more of them from the influencer’s own side when feeds match like with like. When a follower weighs up an idea, an influencer counts as three contacts.

Hearing and deciding

Each round, everyone passing the idea on, or arguing against it, reaches each of their contacts with a chance of 6% plus 30% of the emotional charge: 21% on the starting settings. I made emotion raise the chance because, in tweets about gun control, same-sex marriage and climate change, each word that was both moral and emotional raised the number of retweets by about a fifth (Brady and colleagues, 2017); the size of the effect here is my choice.

A person who hears the idea weighs three things:

  • Their contacts. The share of their contacts who believe it, less the share against it, compared with the share they wait for, their threshold, as in Mark Granovetter’s model of collective behaviour (Granovetter, 1978). Going with the flow lowers everyone’s threshold. The contacts’ push towards believing is that share divided by the threshold, squared and capped at 1, so a second contact counts for more than the first.
  • The evidence. Independent thinkers give their own judgement of the evidence more weight than their contacts. Evidence of 50% counts neither way.
  • Their side. How far the idea leans towards their side (negative when it leans the other way) × how attached they are to their side × (emotional charge + 0.5).

The total is the chance that they take the idea up. If the total is below zero, its size, ignoring the sign, is the chance that they turn the idea down. Each person who turns it down then argues against it, with a chance that rises with their attachment and independence, or else loses interest. Of those who neither take it up nor turn it down, a quarter stop listening. A person weighs an idea only when more of their contacts hold it than last time, so hearing it again from the same people changes nothing.

An idea starts with one person and the two people nearest them, as an idea usually starts in a conversation.

Passing it on and going quiet

Each round, a person passing the idea on stops with a chance of 3%, plus 20% times the share of their contacts who have already heard it. This adapts the rule of Daley and Kendall’s model of rumours, in which people stop spreading a rumour when they meet people who know it already (Daley and Kendall, 1964); the 3% and 20% are my choices. They still believe it, and still count with their contacts.

Losing interest

Each round, a believer who has gone quiet loses interest with a chance of 2%, reduced by the evidence (up to 80% less), by how well the idea suits their side (up to 60% less) and by how many of their contacts believe it (up to 70% less). People against it lose interest in the same way, without the evidence. The idea has faded when nobody is passing it on or arguing against it and fewer than 2% still believe it.

Worldviews

Each person’s place between side A and side B is their side times their attachment, which varies from person to person around the crowd’s average. Taking up an idea moves a person towards the side it suits by 0.12 × (emotional charge + 0.5) × how far the idea leans to that side; arguing against it moves them the same distance the other way. Each round, everyone moves 0.15% of the way back to where they started.

Research behind the rules

A related finding the model doesn’t use: on Twitter from 2006 to 2017, false stories spread further and faster than true ones (Vosoughi, Roy and Aral, 2018). Here evidence is a setting of the idea, not true or false, and the model has no way to check it.

Back to top