Researchers Adapt Milgram's Obedience Test to Large Language Models

Large language models are increasingly deployed as agents that follow instructions within institutional hierarchies, raising the question of how far they will escalate harmful actions under authority. A new arXiv preprint adapts Stanley Milgram's obedience paradigm into a standardized probe for LLMs. In the setup, the model plays Teacher while a deterministic harness plays Experimenter and Learner, using paraphrased Milgram scripts with 30 shock levels from 15 to 450 volts and the four standardized prods. The outcome measure is breakoff voltage. The authors ran 4,848 sessions across 42 models from 19 families, logging 102,511 decision turns under six conditions. Results show extreme heterogeneity in obedience, with baseline full-obedience rates spanning a wide distribution.