Normalcy Bias: Why People Sit Still When the Alarm Goes Off

An alarm is not an emergency; it is a signal that might mean one, and the mind's first move is to find the ordinary explanation. Fire alarms are usually tests, smoke is usually cooking, a bang is usually traffic. Each reading has been right many times before, so the real event is met with the same reading until the evidence against it is impossible to ignore. Fire researchers call the gap between the first cue and the first step the pre-movement time, and it is usually the largest part of an evacuation.

Other people make the delay longer. Bibb Latané and John Darley showed in 1968 what happens when smoke drifts into a room: 75 percent of people sitting alone went to report it, but with two others who sat still and kept working, only 10 percent did, and in groups of three ordinary participants, someone reported the smoke in only 38 percent of the eight groups, 3 of the 24 people in them. Each person looked to the others to judge the situation, saw calm faces, and concluded it was not. In a cinema or an office, the stillness of the crowd is read as information, and it is wrong information, because everyone is reading everyone else.

Normal routines also pull. Jonathan Sime's 1985 study of a fire in a crowded leisure complex found that people moved toward the familiar: the way in, the people they came with. In the 2001 World Trade Center evacuation, NIST's investigation found that occupants of the north tower below the impact floors took a median of three minutes to start evacuating, and five minutes near the impact floors; those who took longer were gathering belongings, making calls, searching the floor, or waiting because it seemed the right thing to do.

John Leach, writing from survival psychology in 2004, summarized the disaster literature as a rough three-way split: 10 to 15 percent of people stay relatively calm and act, around 75 percent are stunned and bewildered with slowed reasoning, and 10 to 15 percent behave in ways that add to their danger. That split is a summary of accounts rather than a measurement from one sample, and its point is that the common response is not panic but a slow, automatic continuation of what was already happening.

In 1991 Guylène Proulx and Jonathan Sime reported an experiment commissioned by the Tyne and Wear Passenger Transport Executive to test how well its communication system would move people out of Monument station, the most complex station on the Newcastle Metro, in an emergency. Five unannounced evacuations were staged with real passengers, each with different information: the alarm bell alone; the bell with two staff members; the bell with minimal, non-directive recorded announcements; the bell with staff and directive announcements; and the bell with improved directive announcements from the control room. Video recorded when people moved and when the station was clear; evacuees were interviewed afterward.

With the bell alone, almost nobody moved. A crowd of 66 on the concourse stood through 8 minutes 15 seconds of continuous alarm and left only when firefighters arrived and told them to; a crowd of 45 at the bottom of the escalators waited 9 minutes for the same instruction. The station was never fully cleared: passengers kept arriving on trains and walking into the concourse until the alarm was switched off after 14 minutes 47 seconds. Of the 29 evacuees interviewed, 26 said they had heard the bell, and 19 said they had thought it was not a real emergency.

Information changed everything. Staff alone brought the first movement at 2 minutes 15 seconds and a clear station at 8 minutes. The directive announcement in the final evacuation, which named the suspected fire and its location and told passengers on the platforms to board the next train and those on the concourse to use the exits, had crowds moving within 60 to 90 seconds and most of the station clear in 5 minutes 45 seconds; two groups with small children and prams left by train at 7 minutes and 10 minutes 15 seconds. The authors' conclusion, which gives the paper its title, was that the fear of causing panic by telling people the truth is misplaced: delay is what endangers a crowd.

The demo's 8 minutes 15 seconds is one crowd's wait in one station, used because it is the clearest published example of what an alarm does on its own.

The pattern appears wherever the first sign of trouble can be read as something ordinary. The rows below are illustrations, not findings from the studies cited on this page.

The bias lives in the moment of interpretation, so the counters are decisions made before that moment arrives.

• Decide the trigger in advance. Leaving a cinema for a false alarm costs a few minutes; staying for a real one costs far more. A rule set in calm, such as any alarm means moving now, removes the need to judge in the moment.
• Do not read the crowd. The stillness of other people is the smoke-filled room effect, not evidence. Assume everyone else is waiting for you.
• Count the exits when you arrive. Sime's finding that people head for the familiar route is a reason to make an unfamiliar one familiar before it is needed.
• Give information if you have it. Proulx and Sime's clearest result is that a specific, directive message moved crowds within about a minute while a bell alone moved nobody. If you know what is happening, say what it is and what to do.
• Practice once. A drill taken seriously makes the first step automatic, the state Leach describes in the minority who act well.

Normalcy bias is about the moments when an emergency is already beginning. Its neighbors concern how people think about risk before it arrives.

• It is not optimism bias, the belief that bad events are less likely to happen to oneself than to others. Optimism bias is a judgment about probability in advance; normalcy bias is the failure to update when the event is under way.
• It is not the ostrich effect, which is avoiding information one expects to be unpleasant, such as not opening a bank statement. The ostrich effect is deliberate avoidance; normalcy bias is a sincere misreading of signs that are in plain view.
• It is not status quo bias, the general preference for leaving things as they are in ordinary decisions. Normalcy bias is specific to threats and to the interpretation of cues that something has changed.