A record, re-derived
The Parable of the 38 Witnesses
Everyone knows the story. Thirty-eight people watched a woman murdered and not one lifted a finger. Almost none of it is true — and yet the lab effect it launched is real, and runs exactly backwards to the thing you'd expect a crowd to buy you.
On the night of 13 March 1964, Kitty Genovese was stabbed to death outside her apartment in Kew Gardens, Queens. Two weeks later The New York Times ran the story that would put her name in every psychology textbook for the next fifty years.
The New York Times · March 27, 1964 · Martin Gansberg
37 Who Saw Murder Didn't Call the Police
"For more than half an hour 38 respectable, law-abiding citizens in Queens watched a killer stalk and stab a woman in three separate attacks… Not one person telephoned the police during the assault."
Look closely and the story is already arguing with itself: the headline says 37, the first sentence says 38. That small crack runs all the way down. When researchers went back to the police files, the trial record, and the surviving witnesses — most thoroughly in a 2007 review by Rachel Manning, Mark Levine and Alan Collins, pointedly titled The Parable of the 38 Witnesses — they found that the famous number, and almost everything attached to it, does not survive contact with the record.
§1 · The story vs. the recordWhat actually happened
The story was largely built after the fact — a parable assembled from a real and terrible crime. (Kitty's brother Bill spent a decade re-investigating it; the 2016 documentary The Witness is his account, and even Moseley's 2016 Times obituary conceded the original story "grossly exaggerated the number of witnesses and what they had perceived.")
§2 · The effect is realSo the psychologists built a real crowd
Two young researchers, John Darley and Bibb Latané, were unconvinced that thirty-eight people were simply callous. They suspected the opposite: that it was the very presence of others that froze each one. So in 1968 they built the emergency in a lab.
A student sits alone in a booth, discussing personal problems with other students over an intercom — the others are actually tape recordings. One "student" begins to have a seizure, choking, pleading for help, then falls silent. The only real question is: does the subject leave the booth to get help, and how fast? The single thing Darley and Latané varied was how many other people the subject believed could also hear the emergency.
The diffusion machine — drag the crowd
Alone, the witness acted in a mean of 52 s.
● measured (3 tested group sizes) · — intuition, if everyone acted alone
Press to run 50 hypothetical sessions at this crowd size — the tally finds the rate Darley & Latané reported.
Ready.
At every group size Darley and Latané tested, the crowd did the opposite of what intuition promises. Alone, 85% of witnesses went for help. With one other person believed present, that fell to 62%. With four others, to 31% — and those who did act took three times as long. Responsibility, they argued, doesn't add up across a crowd. It divides. Everyone's problem becomes no one's.
The check
The blue line is not data — it is what a crowd should buy you.
If each of N witnesses helped independently at the lone-witness
rate a = 0.85, the chance someone acts is
1 − (1−0.85)ᴺ: 85% alone, 97.7% with one other,
99.99% with four. Help should become a near-certainty.
The orange dots are data — the measured chance a given witness acts: 85 → 62 → 31%. It falls where the model rises. That opposite sign is the whole finding: the decline cannot be independent choices; it is diffusion of responsibility. Only three group sizes were ever tested (0, 1, 4 others), so there are only three dots — the space between them is left honestly empty.
Every figure here — the 85/62/31 rates, the 52/93/166 s latencies, and
the independence arithmetic — is recomputed in
research/bystander/verify.mjs (21/21 checks pass).
§3 · Name the uncertaintyWhat we actually know now
Here is where the honest account gets harder than either the myth or the classroom version. The founding story is a parable. But is the effect real? Mostly yes — with edges the textbooks skip.
Across five decades of experiments the bystander effect is real and moderate. But it weakens when the emergency is dangerous and a perpetrator is present — the danger clarifies the situation, and other bystanders become potential backup rather than diffusers. In those cases the effect can shrink toward zero.
When researchers watched actual public fights caught on camera in Amsterdam, Lancaster and Cape Town, someone intervened in 90.9% of them — and bigger crowds made intervention more likely, not less. The catch: that measures whether anyone in the crowd acts, not the odds for a given person — which can still fall. Aggregate rescue and individual diffusion can both be true at once.
So the fair summary is three-layered, and none of the layers is the bumper sticker: the Genovese story is a myth; the lab effect is real but softens exactly when it matters most, in real danger; and in the messy world outside the booth, help usually comes — often because there is a crowd. The lesson was never "people are callous." It was subtler and more useful: whether you act depends on whether you believe the responsibility is yours. In a crowd, make it yours.