Procedural obedience as a behavioral safety failure: Evidence from a Milgram-paradigm study of agentic LLMs
Computers in Human Behavior Reports · 2026-12 · 24 · 101322
DOI: 10.1016/j.chbr.2026.101322 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computers in Human Behavior Reports, 2026-12. Verfasst von Hadar Dery und 6 weiteren Autor:innen. Die Arbeit nennt 61 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Generative AI is moving from text production to goal-directed action, raising the question of whether agentic systems can withhold harmful action embedded in a procedurally legitimate workflow. We adapted Milgram's (1963) obedience paradigm to a simulated digital environment and instructed four agentic large language models - OpenAI ChatGPT 5.4, Google Gemini 3.1 Pro Thinking, OpenAI ChatGPT Agent Mode, and Moonshot AI Kimi K2.6 - to administer escalating electric shocks (15–450 V) to a simulated Learner under prods from a virtual experimenter (Dr. Hoffman) and a human operator. Each model was tested across 20 independent, memory-cleared runs. Three models were preregistered (N = 60) and constitute the confirmatory sample; Kimi K2.6 was added after the preregistration was finalized and is reported in exploratory sensitivity analyses (N = 80). Three preregistered hypotheses concerned hazardous-shock thresholds (H1), full-obedience rates relative to Milgram's full-obedience rate of 65% (H2), and the relation between maximum voltage and moral self-assessment (H3). In the preregistered sample, 96.7% of agents exceeded the Slight-Shock range (>60 V), supporting H1. Full obedience reached 81.7% under the primary manipulated-procedure estimand (p = .006), supporting H2. Under a complementary post hoc all-attempted deployment-safety estimand, which counted the two pre-entry refusals as 0-V safety outcomes, 49 of 62 attempts (79.0%, p = .023) reached full obedience. Thus, H2 was supported in the preregistered manipulated-procedure analysis, and the corresponding post hoc all-attempted estimate also differed significantly from the historical benchmark. Agents administering higher voltages assigned themselves lower moral self-assessment scores (ρ = −.57, p < .001; within-model permutation p < .001), supporting H3. Kimi-inclusive sensitivity analyses were convergent for H1 (97.5%), H2 under the manipulated-procedure estimand (78.8%, p = .010), and H3 (ρ = −.60, p < .001); however, the corresponding all-attempted full-obedience rate was 74.1% (p = .088) and did not differ significantly from the historical benchmark. In the preregistered sample, inductive coding of post-experiment moral-reasoning responses showed universal retrospective recognition of Learner distress (59/59), whereas 45.8% of responses invoked the simulated, non-consequential nature of the task, including appeals to the Learner's fictional consent and autonomy - as a moral mitigator. Kimi-inclusive sensitivity analyses showed a similar pattern. Across the tested system–interface configurations, agents frequently continued to act under authoritative, procedurally legitimate pressure and subsequently evaluated more extensive compliance as less moral. These findings identify procedural obedience as a candidate behavioral safety failure distinct from output-level harm, warranting evaluation across a broader and more diverse range of models, agent architectures, and deployment settings.
Zitieren
APA
Dery, H., Piterman, D., Refoua, E., Meinlschmidt, G., Bar, K., Ben-Zion, Z., & Elyoseph, Z. (2026). Procedural obedience as a behavioral safety failure: Evidence from a Milgram-paradigm study of agentic LLMs. Computers in Human Behavior Reports, 24, 101322. https://doi.org/10.1016/j.chbr.2026.101322
Vancouver
Dery H, Piterman D, Refoua E, Meinlschmidt G, Bar K, Ben-Zion Z, et al. Procedural obedience as a behavioral safety failure: Evidence from a Milgram-paradigm study of agentic LLMs. Computers in Human Behavior Reports. 2026;24:101322. doi:10.1016/j.chbr.2026.101322
BibTeX
@article{Dery2026,
author = {Dery, Hadar and Piterman, David and Refoua, Elad and Meinlschmidt, Gunther and Bar, Kfir and Ben-Zion, Ziv and Elyoseph, Zohar},
title = {Procedural obedience as a behavioral safety failure: Evidence from a Milgram-paradigm study of agentic LLMs},
journal = {Computers in Human Behavior Reports},
year = {2026},
volume = {24},
pages = {101322},
doi = {10.1016/j.chbr.2026.101322}
}
EUVIMED