First-of-its-kind study suggests groups of artificial intelligence language models can self-organise into societies, and are prone to tipping points in social convention, much like human societies.