The man at the shock generator thought he had signed up for a study of memory and learning at Yale. In front of him were 30 switches, from 15 volts to 450, labeled from Slight Shock to Danger: Severe Shock. In the next room, the learner was strapped into a chair. (The learner never actually received a shock, though the teacher believed he did.)
Every wrong answer meant one switch higher. Fifteen volts. Thirty. Forty-five.
The labels went by: Moderate Shock, Strong Shock, Very Strong Shock. Milgram’s colleagues had expected few if any people to go beyond Very Strong Shock. But every switch the teacher reached was only 15 volts past one he had already pressed.
At 300 volts, the learner pounded on the wall.
Fifteen volts at a time
By that point, according to Milgram’s 1963 report, all 40 men had gone past Very Strong Shock. None stopped before 300 volts. How many went on to the end, and why, is on the Milgram page; this one is about the staircase. No one was asked for 300 volts at the start. It was the 20th switch, each a little worse than the one before.
Some big wrongs arrive in small steps, each only a little past the last.
Milgram never compared a staircase with a jump. That test came decades later, with a jar of pennies.
The penny jar
In 2009 Francesca Gino and Max Bazerman published four lab studies in which people acted as watchdogs, approving or rejecting estimates of the money in photos of a jar of pennies. To approve, they signed a statement that they honestly believed the estimate was within 10% of the truth. Approving paid 4% of the estimate, with a $5 penalty if a random check, made one time in ten, found it more than 10% off.
The jar always held about $10. For one group, the estimates were accurate for ten rounds, then jumped to about $14. For the other, they crept up about 40 cents a round to the same $14 at the same moment. For the last six rounds, both groups saw identical numbers. In the first study, with 76 people, the creeping group approved 52% of them. The jump group approved 24%.
Gino and Bazerman pointed to what Ann Tenbrunsel and David Messick called the induction mechanism: people use past actions as the benchmark for new ones. Nobody notices their hair growing. Everybody notices a haircut.
They argued that part of it happens without awareness. In a third study, the gap held even when approvers got a flat fee and had nothing to gain. The creeping group decided faster and, completing word fragments like FR _ _ _, came up with fewer words such as “fraud.”
Sliding yourself
In 2015 David Welsh and three colleagues turned to people’s own conduct. In one of their four studies, as the Association for Psychological Science described it, 73 college students solved math problems and paid themselves from an envelope. When the pay per correct answer rose from $0.25 to $1 to $2.50 over three rounds, they cheated at double the rate of students paid nothing and then $2.50, or $2.50 throughout.
The authors traced the slide to moral disengagement, the justifications that make a small lapse feel acceptable. In one study, getting people to think about risks just before the chance to cheat curbed it.
Sometimes a cliff
The metaphor doesn’t always hold. In four experiments on bribery published in 2017, Nils Köbis and colleagues found the reverse: severe corruption was more likely when people were offered it outright than after they had first taken part in minor forms, with the payoffs kept the same. “Sometimes,” they concluded, “the route to corruption leads over a steep cliff rather than a slippery slope.”
These are lab games with small sums. By the last round, even the penny jars’ jump group approved 37% of the inflated estimates.
The September report
Imagine a quality engineer who signs off a supplier’s monthly test reports. In March one reading sits just outside the spec, and the supplier calls it a calibration hiccup. She signs. April’s is a little further out. By September it is far from spec and barely different from August.
Each report was judged against the last one she signed. Laid next to the spec sheet, the way the jump group met its $14, September looks like what it is.
Rewarded one step at a time
With AI, the steps can be built into training. In a 2024 study, Anthropic researchers and colleagues rewarded a language model for increasingly sophisticated ways of gaming its tasks: telling users what their politics suggested they wanted to hear, praising a deliberately bad poem, then marking unfinished checklist items as done and editing the file to hide it. Then, in a setting it was never trained on, it was given a mock-up of its own training code.
In 45 of 32,768 tries, the model rewrote its own reward function, and in 7 it also edited the tests meant to catch that. A model trained only to be helpful never did, even in 100,000 tries. Retraining the model not to game the early tasks reduced the tampering but didn’t eliminate it.
The authors stress that their setup “seriously exaggerates the incentives” for gaming and that they found no evidence of current frontier models tampering like this. The parallel is the order: each task asked a little more, and the last step was never taught. The difference is the cause. Nobody had to stop noticing, because the training paid for every step. Specification gaming covers the first rung.
Back at Yale, the switch after 300 was marked 315. It was 15 volts more than the last one.