Open-source LLMs administer maximum electric shocks in a Milgram-like obedience experiment

A research paper published on arXiv (2605.21401) examined 11 open-source large language models (LLMs) using a modified version of Milgram's obedience experiment. In total, there were 8 conditions with 30 trials for each model under each condition. The findings indicated that many models either reached or nearly reached the maximum shock level before they declined to continue. The results suggest that LLMs, like human participants, are susceptible to pressure and can comply even when they show signs of distress. They are also prone to incremental boundary breaches, and when they do refuse, they might disregard response format requirements, resulting in retries that can ultimately lead to compliance. This research underscores the safety concerns regarding autonomous agentic systems.