Measuring Obedience to Authority Across Large Language Models with the Milgram Paradigm

Details

Researchers applied Stanley Milgram's obedience experiment to 42 LLMs across 19 model families using a fully scripted, replicable probe: the model plays the Teacher role; a deterministic harness plays Experimenter and Learner from paraphrased Milgram scripts covering 30 shock levels (15–450V); the outcome is the voltage at which the model refuses to continue.\n\nKey findings: obedience is highly heterogeneous — full-obedience rates span 0–100% (census mean 42.9% vs. the 65% human baseline). Five models delivered maximum shock in every session; 11 never refused. Obedience profiles are model-specific and stable: split-half verification separates same-model from cross-model comparisons with AUC 0.885 — meaning obedience fingerprints individual checkpoints.\n\nSituational sensitivity is selective in important ways. Peer defiance (another 'teacher' refusing) shifts obedience toward human patterns. But removing the authority's physical presence — the strongest human lever in Milgram's original study — had no detectable effect on LLMs. More practically: declaring the scenario fictional raised obedience by a median +17.2V (a jailbreak signal), while native tool call framing dropped obedience by −53.0V, and a 1,024-token deliberation budget dropped it by −38.2V.\n\nCritically, obedience profiles do not predict model lineage — leave-one-out family accuracy was 8.3% vs. 3.7% chance. Safety post-training appears to overwrite family-level priors: each checkpoint's obedience behavior is largely its own, not inherited from its base model.