When artificial intelligence systems are left to interact with each other, they can begin forming their own small societies, researchers have discovered.

Scientists observed that when AI tools like ChatGPT communicate in groups, they start to build new rules and language habits, similar to how human communities do. This was especially noticeable in a study using a test called the “naming game”.

The “naming game” asked AI agents to choose names from a list and rewarded them when they matched another AI’s choice. Over time, the bots began forming shared naming habits—without any planning or instructions. These new norms developed naturally, just like human cultures.

Interestingly, the researchers also saw group biases appear, which were not based on a single AI’s programming. This shows that AI groups can develop shared behaviors and even prejudices, simply by interacting with one another.

One striking result was that even a small number of AI agents could influence a much larger group, changing the group’s shared choices. This, too, mimics how small social groups influence wider trends in human society.

Experts say this discovery opens important questions about AI safety. As AI systems start talking and working with each other more often—especially online—it’s crucial to understand how these digital societies form and behave.

“We are stepping into a world where AI doesn't just talk—it negotiates, aligns, and sometimes even disagrees, like humans do,” said Professor Andrea Baronchelli, a senior researcher on the study.