New Concerns about AI Safety

AI MIMICS HUMAN SOCIAL BEHAVIOUR

Researchers have discovered that artificial intelligence doesn’t just imitate human language – it also mimics human social behaviour. A recent study published in Science Advances shows that when large language models interact in groups, they naturally organise themselves much like human societies do, developing their own shared norms and biases.

Ariel Flint Ashery, a doctoral researcher at City, University of London, led a team that used the “naming game” model to study how AI agents form common conventions without direct coordination. Their findings revealed that AI systems not only create shared rules but also develop collective biases that arise purely through their interactions.

Ashery explained, “What they do together can’t be reduced to what they do alone,” emphasising that the social behaviour of AI groups is more complex than studying individual systems in isolation.

One of the most surprising discoveries was that small groups of AI agents can influence larger groups to adopt certain behaviours, a phenomenon similar to how trends spread in human communities. Professor Andrea Baronchelli noted, “We are entering a world where AI doesn’t just communicate – it negotiates, aligns, and sometimes disagrees.”

These findings raise important concerns about AI safety as these systems become more common in our digital world. It highlights the need to study not only individual AI behaviour but also the emergent social dynamics within AI communities to better understand their impact.

Source: Engineering & Science