The person who steps in gets hit 3.6 percent of the time

Of 337 bystanders who only de-escalated real street conflicts on CCTV, 12 were hit. No weapons, no head-butts. What that number does and does not mean.

Everyone has heard the story about the thirty-eight witnesses who watched Kitty Genovese die and did nothing. It has been in psychology textbooks since 1968 and in newspaper columns ever since, usually offered as proof that cities make people cold.

The number was wrong. But the more interesting correction is not to the number.

The aggregate question

Half a century of bystander research asked whether an individual is less likely to help when other people are present. Broadly, yes. What nobody could answer, because answering it meant watching real emergencies instead of staging them in a lab, was the question the person on the pavement is actually asking: will somebody help me.

Richard Philpot and four colleagues went and looked. They obtained CCTV footage of public disputes from city-centre cameras in Amsterdam, Cape Town and Lancaster, and coded 219 aggressive incidents. In nine out of ten of them, at least one bystander intervened, and help usually came from more than one person. That work ran in American Psychologist in January 2020.

The lab result was not wrong. It was answering a different question, and the different question is not the one that matters when you are the one on the ground.

The question nobody had asked

Then comes the objection any sensible person raises next, and it is the reason a lot of us hesitate. Fine. Somebody usually helps. What happens to them?

That paper exists too, and it is the more useful one. Lasse Suonperä Liebst, Philpot, Mark Levine and Marie Rosenkrantz Lindegaard published it in Psychology of Violence in 2020, working from the same footage archive. Out of 1,225 raw clips they kept 219 that met their criteria — inner-city, at least two antagonists, image quality good enough to code behaviour frame by frame. Coding video by hand is slow, so they took a random 93 of those and coded every intervening bystander in them: 417 people. As a comparison group they coded 636 bystanders who were present and stayed out of it.

From the 417 they isolated 337 who used only de-escalatory behaviour — no shoving, no punches of their own, nothing a camera could read as joining in. Twelve of those 337 were physically harmed. That is 3.6 percent.

Among the 602 uninvolved bystanders the model could compare them against, exactly one person was harmed. Zero point two percent.

So intervening does raise your risk. Significantly, and by a factor of about eighteen. It raises it from almost nothing to nearly nothing.

What the harm actually was

This is the part that changed how I read the number. Of the twelve, six were a single blow and five were shoving. There were no head-butts. Nobody was attacked while on the ground. No weapon appeared in any of the ninety-three clips.

The picture most of us carry when we imagine stepping into a street fight — the escalation, the thing pulled out of a pocket, the ambulance — is not in the data. What is in the data is a shove.

Two things that did not predict risk

The researchers expected that intervening in a violent conflict would be more dangerous than intervening in a shouting match. It was not. The odds ratios sat around one, and the Bayes factor came out at 4.79 in favour of no relationship at all: the data were roughly five times more consistent with severity being irrelevant than with it mattering.

The number of bystanders present did not predict risk either.

What did predict it was affiliation. If you were connected to one of the people fighting — a friend, a partner, someone from the same group — your odds of being hit were about five and a half times higher. In absolute terms the model put it at 5 percent for the affiliated against 2 percent for everyone else, and the authors add that the 2 percent cannot be told apart from zero.

Which means the stranger who walks over and the friend who was already standing there do not face the same situation. In news reports they are one story. In the data they are two.

Where I would slow down

Gender showed up as a predictor, with men more often harmed, odds ratio 3.82. But its Bayes factor was 1.54, and the authors say plainly that at that value the data cannot distinguish the hypothesis from its absence. A finding on paper, not much more. Risk also varied by city: higher in Lancaster than in Amsterdam, with Cape Town not distinguishable from either. Three cities and ninety-three clips is no basis for saying anything about countries, and the paper does not try.

The whole study describes one narrow situation — a public dispute, in a city centre, on camera, where the person stepping in is de-escalating. Behind a closed door it says nothing at all. Nor about a knife already visible, or about someone being quietly followed rather than shouted at. And 3.6 percent is a fact about three hundred and thirty-seven people. It is not a fact about you, on a Tuesday, outside a particular bar.

Why this sits on a map of good deeds

Almost nothing recorded on Good People Map is a fight. It is a lift to the airport, a drill handed over a fence, someone's cat found at two in the morning by a neighbour who did not have to go outside.

The connection is not the violence. It is the arithmetic that runs underneath all of it, in the two seconds before somebody decides — the silent estimate of what this will cost me. That estimate is nearly always too high. Worth knowing that it stays too high even in the version with the shouting.

Sources: Philpot, R., Liebst, L. S., Levine, M., Bernasco, W., & Lindegaard, M. R. (2020). Would I be helped? Cross-national CCTV footage shows that intervention is the norm in public conflicts. American Psychologist, 75(1), 66–75. Liebst, L. S., Philpot, R., Levine, M., & Lindegaard, M. R. (2020). Cross-national CCTV footage shows low victimization risk for bystander interveners in public conflicts. Psychology of Violence, doi:10.1037/vio0000299.