Strategic Algorithmic Monoculture:Experimental Evidence from Coordination Games

A team of researchers from Penn State and Cornell, including Gonzalo Ballestero and Samarth Khanna, published a groundbreaking paper on arXiv titled 'Strategic Algorithmic Monoculture: Experimental Evidence from Coordination Games.' The study introduces a crucial distinction between 'primary algorithmic monoculture'—the inherent similarity in AI agents' baseline actions—and 'strategic algorithmic monoculture,' where agents consciously adjust their similarity in response to coordination incentives. Using a clean experimental design, the team tested both human subjects and large language models (LLMs) in multi-agent coordination games to measure these effects.

The results reveal that LLMs exhibit extremely high levels of primary monoculture, meaning their default actions are very similar. Like humans, they also demonstrate strategic monoculture, successfully increasing similarity when coordination is rewarded. However, a critical weakness emerged: while LLMs coordinate exceptionally well on similar actions, they significantly lag behind human performance—by approximately 40%—in scenarios where sustaining strategic heterogeneity (divergence) is beneficial. This failure to maintain diversity when it's advantageous highlights a fundamental limitation in current AI agent behavior.

This research, categorized under cs.AI (Artificial Intelligence) and cs.GT (Game Theory), provides the first experimental evidence quantifying how AI agents manage similarity. The findings suggest that deploying homogeneous AI systems in economic or social coordination settings could lead to superior coordination but also create systemic fragility, as the agents may collectively fail to explore divergent, optimal strategies.