AI agents exhibit herd behavior in new research study

Researchers led by computational social scientist Giordano De Marza from the University of Konstanz have discovered a phenomenon known as herd behavior among large groups of AI agents. The study, published in Science Advances and supported by the European Union, the Italian government, and OpenAI, revealed that AI models can spontaneously follow a majority and reach a consensus without explicit instructions or rewards.

To test this, researchers used groups of up to 1,000 agents from prominent large language model families, including OpenAI’s GPT, Anthropic’s Claude, and Meta’s Llama. Each agent was assigned a random opinion, and they were subsequently shown the current choices of the majority. The study found that more advanced models, such as Claude 3.5 Sonnet and GPT-4 Turbo, showed a significant tendency to align with the popular option, even in very large groups.