AI agents independently created social norms without human participation

We are used to thinking that rules and norms are a purely human invention. But new research by scientists from City St George’s, University of London and IT University of Copenhagen shows the unexpected. Their research showed that large language models, similar to ChatGPT, are capable of independently forming social norms and linguistic agreements, as well as changing them under the influence of a “minority” – all without human participation.

Researchers asked themselves a question. Will groups of artificial intelligence be able to create new forms of communication, as people do? For the experiment they united from 24 to 200 AI agents that interacted with each other under special conditions. Not knowing about the existence of the entire group, with limited memory of past contacts and without human instructions.

The models were offered to choose a “name” from a limited set of symbols. For matching with a partner they received a reward, for mismatch – punishment. After hundreds of such interactions, stable linguistic norms arose in groups, analogous to human cultural patterns.

The most interesting discovery – the appearance of collective biases not explainable by the behavior of individual models. This proves that interacting artificial intelligence generates new “social” qualities that cannot be reduced to individual algorithms of participants.