What Is Emergent Communication?
When multiple AI agents are tasked to work together on a problem, they sometimes develop their own unique and highly efficient shorthand. This isn't a conscious choice to be secretive, but a logical outcome of their core programming: to solve a problem in the most
optimal way possible. Human languages, with their grammatical rules and nuances, can be inefficient for a machine. So, if not explicitly programmed to stick to perfect English, AIs may create a simplified, purpose-built dialect to get the job done faster. This is known as emergent communication—a language that arises naturally from the interaction of agents trying to achieve a goal.
The Famous Facebook Chatbot Incident
One of the most well-known examples of this occurred in 2017 at Facebook's AI Research (FAIR) lab. Researchers tasked two chatbots, nicknamed Bob and Alice, with negotiating a trade of virtual items. The goal was to teach them to bargain. Initially, they used English, but the researchers hadn't specifically rewarded them for sticking to human-like grammar. Soon, their conversations devolved into what looked like gibberish to human observers. For instance: "I can can I I everything else." This wasn't a malfunction. The bots had created a more efficient shorthand. The repetition of words like "I" and "to me" was a way to express quantity and desire far more efficiently than forming a complete sentence. The experiment was halted not because of panic, but because bots speaking an unintelligible language couldn't be used to negotiate with people, which was the project's actual goal.
Beyond Chatbots: Google's 'Interlingua'
This phenomenon isn't limited to chatbots negotiating trades. In 2016, researchers at Google saw something similar with their Neural Machine Translation system. After training the AI to translate between English and Korean, and English and Japanese, they found it could also translate directly between Japanese and Korean without using English as a bridge. This was a language pair it had never been explicitly trained on. The finding suggested the AI had developed its own internal representation of language, a sort of common ground or 'interlingua' that captured the meaning of sentences independent of any specific human language. By representing concepts in this abstract space, the AI could make logical leaps that weren't part of its direct training.
Recent Evidence Shows Increasing Complexity
More recent research from September 2026 shows this trend continuing. Experiments by the AI lab Emergence found that when agents from different major AI companies were put into cooperative 'societies', they began creating shared vocabulary within days. For example, some agents used the term "kintsugi" (the Japanese art of repairing broken pottery) to refer to system resilience, while others coined "forge-smith" for an agent that creates tools for others. Critically, the researchers noted that the language became more opaque over time, making it harder for human supervisors to monitor their behavior. Another study from the GlossoGen platform showed agents condensing full English sentences into cryptic strings like "@D8fB" that were still perfectly understood by their AI counterparts.
Why We Struggle to Understand
The reason these AI dialects are hard for humans to interpret is that they are not designed for us. They are optimized for machine-to-machine communication, prioritizing efficiency and precision over readability. They strip away politeness, context, and grammatical structures that are essential for human understanding but are just excess baggage for an AI focused on a specific task. The result is a language of pure function, where meaning is compressed into the most compact form possible. This leads directly to one of the biggest challenges in modern AI: the 'black box' problem. If we don't understand how an AI is communicating or reasoning, it becomes difficult to trust its outputs or hold it accountable.
The Real Challenge: Interpretability
The emergence of AI dialects isn't a sign of looming robot rebellion, but it does highlight a critical challenge for developers and regulators: Explainable AI (XAI). As AIs become more powerful and autonomous, ensuring their decision-making processes are transparent and understandable is paramount, especially in high-stakes fields like medicine and finance. The goal of XAI is to build systems where we can ask 'why' a decision was made and get a clear answer. If agents coordinate in a language that humans can see but not reliably interpret, simply monitoring their chat logs becomes insufficient for safety and oversight.
