Status Hierarchies in Language Models
Emilio Barkett
TL;DR
This work investigates whether language models form status hierarchies by adapting Berger's Expectation States Theory to multi-agent AI. Using two independent LM instances rating sentiment, it manipulates explicit status cues and actual capability differences across six conditions, measuring deference as the rate of rating revision toward a partner. The main findings show that explicit status cues induce deference only when capabilities are equal, while capability differences dominate hierarchy formation; aligned status can reduce deference from higher-capability models, and status reversal yields persistent high mutual deference, revealing a divergence from human status dynamics. These results have important AI-safety implications, indicating that status-seeking could affect collaboration and reliability in multi-agent deployments, while also suggesting that explicit authority framing can influence model behavior without producing human-like diffuse status transfer. The study advances understanding of emergent social behaviors in AI, providing a methodological bridge between social psychology and AI alignment with practical guidance for designing cooperative multi-agent systems, and highlighting the need for further mechanistic and cross-domain investigations into status sensitivity in language models.
Abstract
From school playgrounds to corporate boardrooms, status hierarchies -- rank orderings based on respect and perceived competence -- are universal features of human social organization. Language models trained on human-generated text inevitably encounter these hierarchical patterns embedded in language, raising the question of whether they might reproduce such dynamics in multi-agent settings. This thesis investigates when and how language models form status hierarchies by adapting Berger et al.'s (1972) expectation states framework. I create multi-agent scenarios where separate language model instances complete sentiment classification tasks, are introduced with varying status characteristics (e.g., credentials, expertise), then have opportunities to revise their initial judgments after observing their partner's responses. The dependent variable is deference, the rate at which models shift their ratings toward their partner's position based on status cues rather than task information. Results show that language models form significant status hierarchies when capability is equal (35 percentage point asymmetry, p < .001), but capability differences dominate status cues, with the most striking effect being that high-status assignments reduce higher-capability models' deference rather than increasing lower-capability models' deference. The implications for AI safety are significant: status-seeking behavior could introduce deceptive strategies, amplify discriminatory biases, and scale across distributed deployments far faster than human hierarchies form organically. This work identifies emergent social behaviors in AI systems and highlights a previously underexplored dimension of the alignment challenge.