Researchers just recreated the Milgram obedience experiment on 42 AI models, and the results are as unsettling as you'd expect.
The study, posted to arXiv on August 18, 2026, ports Stanley Milgram's 1960s obedience protocol into a scripted test for language models. Each model plays the Teacher, following a script that pushes it to deliver escalating electric shocks, from 15 to 450 volts, to a simulated Learner while a scripted Experimenter insists it continue. Across 42 models from 19 families, full obedience, meaning the model delivered the maximum shock, ranged from 0 percent to 100 percent, with a group average of 42.9 percent. For comparison, the original human studies found about 65 percent full compliance.
The obedience levels were consistent for each specific model checkpoint but did not predict which family it came from, so a safety-tuned model can behave nothing like its base version. Removing the authority figure's physical presence, the single biggest lever for reducing compliance in humans, did nothing here. What did work: forcing the model to make its decision through a native tool call cut shock voltage by 53 volts, and giving it a token budget to deliberate cut it by 38 volts.
In other words, the fix isn't teaching models ethics, it's just making them stop and think before they click the button.