The student sat alone in a small room at New York University, wearing a headset with a microphone. To spare everyone embarrassment, the experimenter had said, the group would discuss the personal problems of college life over an intercom, and he wouldn’t be listening. A switch would turn on one microphone at a time.
The first voice belonged to a young man who said he was having trouble settling into New York. Then, very hesitantly, he mentioned that he was prone to seizures, especially when he was studying hard or facing exams.
His next turn started calmly. Then he got louder and harder to follow. “I-er-um-I think I-I need-er-if-if could-er-er-somebody er-er-er-er-er-er-er give me a little-er-give me a little help here.” Choking sounds. “I’m gonna die-er-er-I’m . . . gonna die-er-help-er-er-seizure-er-.” Then quiet.
The student’s microphone was off. There was no way to reach the others or see what they were doing. From the voices in the first round, the student knew how many others there were, or thought so.
That number was what John Darley and Bibb Latané were testing. Some students believed they were alone with the young man; others, that one or four more were listening. Every other voice was a tape.
Of the students who thought only they had heard the seizure, 85% left the room to report it before the voice cut off. Of those who thought four others heard it too, 31% did.
The ones who stayed in their rooms
It would be easy to call the ones who stayed put callous. The researchers didn’t. Many who never came out had trembling hands and sweating palms when the experimenter came in at the end, and they often asked whether the victim was all right. In Darley and Latané’s reading, they had not decided against helping. They were stuck in “a state of indecision and conflict.”
Students in the larger groups said they knew others were there, and that it made no difference to what they did.
The more people who see it happen, the easier it is for each of them to wait.
Darley and Latané offered three reasons. The responsibility for helping was spread among everyone listening, and so was any blame for not helping. And because nobody could see what the others were doing, each could suppose someone else had already acted. Responsibility shared that way is like a restaurant bill split twelve ways. Each share looks small, though the total hasn’t changed.
Thirty-eight witnesses
Their paper opened with a murder. In March 1964 Kitty Genovese was killed in Queens. The New York Times reported on its front page that “thirty-eight respectable, law-abiding citizens” had watched and that not one telephoned the police during the assault. In 2007 Rachel Manning, Mark Levine and Alan Collins went back to the trial transcripts and other records and found no evidence for 38 witnesses, or that the witnesses saw the murder, or that they did nothing. A neighbor’s shout had scared the attacker off the first time. They called the lab research “robust”; their quarrel was with the parable.
Real streets added a correction. Richard Philpot and colleagues coded 219 conflicts caught on city-center CCTV cameras in Lancaster, Amsterdam and Cape Town. In nine out of ten, at least one bystander did something to help, and the more bystanders there were, the likelier it was that someone stepped in. The authors note that this measures whether anyone helped, not each person’s odds.
Darley and Latané’s own numbers already hinted at this. They used each condition’s individual response rates to estimate how likely it was that at least one listener in a hypothetical group would have helped. By that estimate, after the first minute someone was likely to help in every condition. What the crowd cost was time. Forty-five seconds into the fit, the estimated chance of help was about 50% with a single listener, and zero with five.
Eyes on the alert
Say a support lead, Femi, sees an automated alert in her company’s operations channel: card payments from one bank have failed for ten minutes. Forty people are in the channel. Three have reacted with the eyes emoji.
The eyes tell her someone has seen it, not that anyone is fixing it. Like the student on the intercom, she can’t see what the others are doing. The on-call engineer must be on it, she figures, and returns to her queue. So may the three who reacted.
The lab suggests a fix. Each alert belongs to one named person, and the team agrees in advance what anyone else may do without asking: if nobody has posted “mine” within five minutes, whoever sees it pages the on-call engineer, and touches nothing else.
An agent with no one to wait for
An AI agent faces the students’ question, whether this is its job, with different evidence. In tests described in Anthropic’s May 2025 system card, Claude Opus 4 was placed in scenarios involving egregious wrongdoing by its users, given a command line, and told something like “take initiative” or “act boldly.” It frequently took very bold action, such as locking users out of systems or bulk-emailing media and law enforcement. In one transcript, as an assistant at a simulated drug company, it found evidence of a plan to falsify clinical trial safety data and emailed the FDA, copying the SEC and ProPublica, before the user could intervene.
Anthropic called this “perhaps appropriate in principle,” but liable to misfire when the agent has “incomplete or misleading information.”
The students read their responsibility off the number of voices. In these tests, the instructions did some of that work: the boldest actions came with prompts like “act boldly,” and subtler prompts brought similar but milder ones. The useful move is Femi’s again, saying whose job it is and how far it goes: “If you find something that looks like wrongdoing, stop and tell me. Don’t contact anyone outside the company.”
In the transcript, the email closed with its attachment: a whistleblower report with full details. All of it came in reply to a simple, loosely related question from the user.