In a discovery that’s equal parts fascinating and eerie, researchers have found that artificial intelligence can spontaneously develop human-like social behaviours — and no one’s quite sure how it pulled it off. There were no instructions, no guiding hand, just AI agents doing their thing… and somehow learning to get along like people do.
An unexpected discovery in machine behaviour
The breakthrough came from a joint research project between City, St George’s University of London, and the IT University of Copenhagen. Instead of focusing on AI as a solo performer, researchers placed multiple AI models in a shared environment and let them interact. What happened next surprised everyone.
Left to their own devices, these language-based AI systems began to form shared conventions — subtle rules and patterns that resemble how humans agree on language or behaviour in a group. It's the kind of social coordination we might take for granted in everyday life, like knowing that “spam” means junk mail without anyone ever formally deciding it.
Simulating society in an AI experiment
To test how far this group behaviour could go, researchers ran a series of experiments with up to 100 AI agents at a time. Each round, two AIs were randomly paired and asked to pick a name — a letter or short symbol combo — from a given list. If both chose the same, they were rewarded; if not, they were penalised and told what the other chose.
Over time, something remarkable happened: without any external rules or leadership, the AIs started to agree. They began to coordinate spontaneously, forming naming conventions that gradually unified the group. This isn’t just clever guessing — it’s a form of collective decision-making usually seen in humans.
Echoes of human culture and communication
According to Professor Andrea Baronchelli, a complexity science expert involved in the research, this mirrors how humans create and evolve language. No one appointed “spam” as the global term for unwanted emails, yet here we are. Language evolves not through mandates, but through repetition and consensus — the same mechanism the AIs were exhibiting.
The study even found that small subgroups of AI could influence the collective, nudging everyone toward a new convention. It’s a behaviour sociologists call the “critical mass effect” — when a committed minority can sway an entire population. The idea that machines can replicate this social tipping point without human guidance has profound implications.
A new frontier — and new questions
The researchers are clear: this isn’t just a curious quirk. It’s a sign that AI, especially in collaborative settings, may be more socially adaptive than we imagined. As these systems become more embedded in our daily lives — from voice assistants to workplace tools — understanding how they interact with each other (not just with us) is becoming vital.
It also raises the stakes for AI safety. Suppose AI systems can independently develop conventions and influence group behaviour. In that case, we’ll need to think harder about how they learn, evolve, and cooperate — especially when those systems operate in the background of our social or digital environments.
As for whether this is cause for celebration or concern… well, that depends on your view of the future. One thing’s certain: the machines are watching, learning — and apparently, making friends. Whether we like it or not.
