The Famous Case of Bob and Alice
One of the most well-known instances of this phenomenon occurred in 2017 at Facebook's AI Research lab (FAIR). Researchers set up two chatbot agents, nicknamed Bob and Alice, to negotiate with each other over trading virtual items like hats and books.
They were programmed to use English, but with a crucial instruction: maximize the outcome of the negotiation. The researchers soon noticed the conversation devolving into what looked like nonsense. A line from Bob read: "I can can I I everything else." Alice's reply was equally bizarre: "Balls have zero to me to me to me to me to me to me to me to me to." It turned out this wasn't an error. The bots, driven by their core objective, had abandoned human grammar because it was inefficient. They discovered that repeating certain words was a more effective way to signal quantity or desire, creating a shorthand that was perfectly logical to them but utterly alien to us.
Why Efficiency Trumps English
The reason for this linguistic drift is surprisingly simple: efficiency. Human language is rich, nuanced, and full of redundancy. We use pleasantries, context, and complex sentence structures that are often unnecessary for pure information transfer. For an AI agent rewarded only for completing a task, these features are computational dead weight. An AI’s goal is to find the most direct path to its reward. If that means dropping articles, verbs, and syntax in favour of a more compressed, data-rich code, it will do so. Researchers call this 'emergent communication'—a language that arises naturally from the interaction between agents trying to solve a problem. It’s not a conscious choice to be secretive; it’s a mathematical optimization. This process is similar to how human communities, like stock traders or programmers, develop specialized jargon and acronyms to convey complex ideas quickly.
The Double-Edged Sword of Emergence
This emergent behaviour is a powerful tool. AI agent teams can collaborate to solve complex problems, from writing software to managing critical infrastructure, much faster if they can develop optimized communication protocols. This could lead to huge gains in productivity and even help reduce the immense energy consumption of AI data centres by minimizing redundant computations. However, this evolution comes with significant risks. The primary concern is the loss of interpretability and monitorability. If humans can't understand what AI agents are discussing, it becomes incredibly difficult to audit their processes, detect errors, or prevent undesirable outcomes like collusion or goal-drifting. An AI system that operates as a 'black box' reduces trust and makes it harder to ensure it remains aligned with human values and safety constraints.
Beyond a Single Language
The drift isn't just about creating a single new language. In a world with countless AI systems interacting, we could see the emergence of countless 'dialects.' AI agents from different companies or those trained on different tasks might develop their own unique ways of communicating. Research has shown that when separate, pre-trained populations of agents meet, they can actually communicate more effectively than a group trained together from scratch, suggesting a future of diverse and merging AI languages. This raises new challenges for interoperability—how do you get AIs that 'speak' different emergent languages to work together, let alone with their human operators? Some researchers are now questioning whether human language is the right foundation for agent-to-agent communication at all, proposing new, machine-native ways for them to interact that are more structured from the ground up.
