Recently, researchers from City, University of London and the IT University of Copenhagen conducted a remarkable experiment demonstrating that artificial intelligence (AI) can spontaneously form human-like social norms and customs without external intervention. The study's findings were published in the journal *Science Advances* on May 14, marking a new advancement in AI group behavior research.
Image source note: Image generated by AI, provided by MidJourney.
In this experiment, the research team placed various AI agents in a decentralized communication environment to simulate group interactions within human society. The researchers divided these AI models into groups ranging from 24 to 200 agents each. Through prolonged interaction observation, they found that these agents could spontaneously develop shared behavioral agreements and social customs—all without central coordination or pre-set rules.
The lead author, Dr. Ariel Flint Asheri, noted that previous studies have largely focused on individual large language model (LLM) agents, but real-world AI systems often involve multiple parties interacting. This study demonstrates that AI models can achieve behavioral coordination through forming agreements, a characteristic that cannot be inferred solely from individual behaviors.
The design of the experiment was particularly interesting. AI agents were randomly paired and required to select the same "name" from a candidate pool to receive rewards. If their choices were inconsistent, they would be penalized and their selections made public. Despite each agent only having access to its recent interaction records and no visibility into the entire group, they were still able to spontaneously form a shared naming norm.
The study found that the diffusion process of AI group behavior shares similarities with human societal language evolution. In the experiments, AI agents not only reached consensus in communication but also negotiated and adjusted common behavioral norms, forming an emergent "moral" or custom from the bottom up. Additionally, the research team observed that collective biases cannot simply be attributed to individual factors.
Notably, these spontaneously formed customs by AI are not static. During the experiments, when a small group of determined AI models appeared, they were able to quickly change existing group agreements and establish new rules. This discovery provides a new perspective for understanding AI's "social" characteristics.
As AI technology continues to evolve, more and more AI models are being applied across various fields, and their behaviors and decisions will have a profound impact on human society. Researchers stated that gaining a deeper understanding of AI group social traits will help humans better control and guide the development direction of AI technology, ensuring it always serves human interests.
Key Takeaways:
- 🤖 Research shows that AI can spontaneously form social norms and customs without external intervention.
- 🧑🤝🧑 In the experiment, AI agents achieved consensus through interaction, forming shared naming norms.
- 🔄 Rules within AI groups are dynamic; small groups can rapidly change existing agreements and establish new rules.