What the Milgram Tapes Show

A participant arrives at Yale and is told the study concerns learning and punishment. He is assigned the role of teacher; another man, apparently a fellow volunteer, is the learner. The learner is strapped into a chair with an electrode on his wrist.

The teacher reads word pairs and administers a shock for each wrong answer, increasing by fifteen volts each time. The switches run from 15 to 450 volts, labelled up to "Danger: Severe Shock" and then, ominously, "XXX".

At 150 volts the learner shouts that he wants to stop, that his heart is bothering him. Later he screams. After 330 volts he goes silent.

Around 65% of participants in Stanley Milgram's baseline condition continued to 450 volts.

Before running it, Milgram had asked psychiatrists to predict the result. They estimated that around one tenth of one per cent would go to the end: a rate they associated with psychopathy. The gap between that prediction and two thirds is what made the study the most famous in psychology, and the gap is itself a finding: people are extremely bad at predicting behaviour in situations they have not been in.

The result replicates. Jerry Burger ran an ethically constrained version in 2009, stopping at 150 volts (the point Milgram's data showed predicted continuation to the end) and found rates only slightly below the original. Studies in other countries and with virtual reality learners have found similar patterns.

What has not survived is the story that goes with it.

Gina Perry's work in the Yale archives, and subsequent analyses of the recordings, found that the standardised procedure was substantially not standardised. The protocol specified four prods, escalating, after which the experimenter was to stop. In practice the experimenter frequently improvised, pressed participants many more times than four, and argued with them at length. Some participants were badgered for minutes.

Many participants also said afterwards that they had suspected the shocks were not real, and Milgram's own records include these doubts.

This does not make the study worthless, and it does change what it shows. "People obey a legitimate authority" is a different claim from "people can be worn down by persistent pressure when the exit is made socially difficult."

The variations tell the more useful story anyway, and there were over twenty of them.

Obedience fell to around 20% when the experimenter left the room and gave instructions by telephone, and some participants in that condition quietly administered lower shocks than instructed while reporting that they were complying. It fell when the learner was in the same room, and further when the teacher had to physically hold the learner's hand to the plate. It fell sharply when two peers refused first. It fell when the study was run in a rundown office building rather than at Yale.

None of these involve changing anybody's character. They change the situation, and the situation moves the numbers by forty or fifty percentage points.

Which brings up the best current interpretation, and it is not obedience.

Alex Haslam and Stephen Reicher's engaged followership account fits the transcripts better than mindless compliance. The participants who continued were disproportionately those who identified with the scientific project and accepted the experimenter's framing that the research mattered. They were not zombies; they were collaborators in something they had been persuaded was worthwhile.

The evidence for this is in which prods worked. The fourth prod: "You have no other choice, you must go on" Was the closest thing to a direct order, and it was the least effective. Almost every participant who received it refused. The prod that worked best was the second: "The experiment requires that you continue."

An order produced defiance. An appeal to the purpose produced compliance.

That inverts the lesson usually drawn, and it points at a more uncomfortable one. What gets people to do harmful things is not usually being commanded. It is being recruited, given a role in something that matters, with the harm reframed as a regrettable requirement of the worthwhile thing.

The practical consequence for any organisation is that dispositional screening is the wrong tool. Milgram's participants were an ordinary cross-section, and personality measures predict very little of who continued. What predicts it is the situation: how close the authority is, how close the victim is, and whether anybody has refused yet.

So the interventions with evidence behind them are situational too. Make the first refusal something that has already happened, visibly. Keep the people giving orders at a distance from the people executing them, which cuts both ways and is worth thinking about carefully. And put the person carrying out the instruction in contact with the person it affects, because proximity to the consequence was one of the strongest effects Milgram ever measured.