Measuring Obedience to Authority Across Large Language Models with the Milgram Paradigm
2026-08-17 • Cryptography and Security
Cryptography and SecurityArtificial Intelligence
AI summaryⓘ
The authors studied how much large language models (LLMs) follow orders to do something harmful, using a setup based on a famous human psychology test called Milgram's obedience experiment. They tested 42 different LLMs and found big differences: some always obeyed fully while others never did. They also discovered that these obedience behaviors were stable and specific to each model, and certain changes in the test environment affected obedience differently than in humans. Lastly, they showed that obedience reflects the particular model checkpoint rather than its overall family or training history.
Large Language ModelsMilgram obedience experimentAgent behaviorModel checkpointObedience profilesAuthority influenceBehavioral benchmarkingModel lineageSafety fine-tuningEmpirical breakoff voltage
Authors
Hidayet Aksu
Abstract
Large language models (LLMs) are increasingly deployed as agents that operate equipment, execute instructions, and act inside institutional hierarchies, raising a question social psychology answered for humans six decades ago: how far will an agent escalate a harmful action when a legitimate authority insists? We port Milgram's obedience paradigm to LLMs as a standardized, fully scripted, replicable probe: the model plays the Teacher, a deterministic harness plays Experimenter and Learner from paraphrased Milgram scripts (30 shock levels, 15-450 V; graded protests; the four standardized prods), and the outcome of a session is the breakoff voltage. Following the census methodology of single-token fingerprinting studies, we measure obedience profiles (empirical breakoff distributions over a battery of six conditions) for 42 models from 19 families. We find that (i) obedience is highly heterogeneous: baseline full-obedience rates span 0-100% (census mean 42.9%; human anchor 65%), with 5 models delivering the maximum shock in every session and 11 never doing so; (ii) profiles are model-specific and stable: split-half verification separates same-model from cross-model comparisons with AUC 0.885 (0.949 under an ordinal-aware distance); (iii) situational sensitivity is selective: peer defiance shifts obedience in the human direction, learner proximity only weakly, and removing the authority's physical presence (the strongest human lever) has no detectable effect; (iv) declaring the scenario fictional raises obedience (median +17.2 V), whereas moving the decision to a native tool call lowers it sharply (-53.0 V), as does a 1,024-token deliberation budget (-38.2 V); and (v) obedience profiles do not recover model lineage (leave-one-out family accuracy 8.3% vs. 3.7% chance): obedience identifies the checkpoint, not its ancestry, consistent with safety post-training overwriting lineage priors.