Chatgpt are spontaneously forming social structures
Imagine a world where artificial intelligences without any human interference begin to create their own social norms, develop their own language and establish conventions that arise from coexistence between them, just as they occur in real human communities.
It seems that that is exactly what a new study has just revealed. Researchers from universities in England and Denmark observed something unprecedented, language models with chatgpt are spontaneously forming their own social structures where they only interact with each other, it all started with a simple but profound question. If we put several large language model (LLM) agents to interact with each other without any specific programming, they would be able to create a culture on their own.
The experiment was inspired by a classic of social psychology
To test this, the researchers created a scenario inspired by a classic of social psychology, the so-called naming game. It works like this, two AI agents are put into conversation, they need to choose a name, a word, a sequence of characters, anything, but with one condition, both have to choose exactly the same name, without having agreed before, if they get it right they win a reward, if they miss they receive a penalty and they only know what the other chose after the attempt.
Only with this simple dynamic and a limited memory of previous interactions did the agents begin to adapt, after hundreds of rounds a spontaneous social convention began to emerge. Most of the agents began to adopt the same naming pattern without anyone having programmed that behavior, but the most impressive thing came later, even when the researchers programmed the individual agents to not have any bias, the group as a collective began to present biased behaviors, this means that when interacting with each other the AIs ended up creating collective preferences that did not exist before and no isolated agent, a kind of already emerging AI culture, where the group begins to have its own social characteristics that no individual agent presented.
And what does this mean for the future of AI? Yes, today with a simple game of names the IAS are already beginning to create norms, social patterns, even their own collective tendencies, what will happen when AI models are released to interact freely on a large scale.
