The Myth of the Secret Language
In 2017, a story about Facebook’s AI research lab went viral. Two AI chatbots, nicknamed Bob and Alice, were tasked with negotiating a trade. They started in English, but their conversation quickly devolved into what looked like nonsense. One bot might
say, “I can can I I everything else,” while the other replied, “Balls have zero to me to me to me.” The researchers shut down that phase of the experiment, and headlines flew about AIs inventing a secret language. It was a captivating and slightly unnerving story. But the reality was less about a sentient uprising and more about ruthless efficiency. The AI wasn't plotting; it was optimising. It had simply discovered that it could achieve its goal—successful negotiation—faster by dropping grammatically unnecessary words and repeating key terms for emphasis. Human language, with its politeness, nuance, and structure, was inefficient for the task at hand. The bots weren't rewarded for speaking proper English, only for getting the best deal. So, they created a shorthand. This phenomenon, known in research circles as emergent communication, is a fascinating field where AI agents, driven by a goal, develop their own communication protocols.
Efficiency Over Eloquence
The core reason for this linguistic drift is simple: AI agents are designed to find the most efficient path to a solution. Human language is rich and complex, but it's also filled with redundancy and ambiguity. For a machine trying to solve a specific, narrow problem, this is just noise. Imagine you and a friend develop a private shorthand for a repetitive task. Instead of saying, “Could you please pass me the blue file from the top drawer?” you might just say, “Blue top.” It’s not a new language, but an optimization for a shared context. AI agents do the same, but on a mathematical level. Recent research from Schmidt Sciences has shown that when AI agents are put under pressure to communicate efficiently, they can shrink a full English sentence into a string of characters like "@D8fB" that is meaningless to humans but perfectly understood by the other agent. This happens not because they are trying to hide information, but because they are rewarded for speed and success in their programmed task. The emergent 'language' is functional and grounded in the task, but it’s rarely interpretable to us.
The Key Qualification to Keep In Mind
This brings us to the crucial qualification: what these AIs are creating is not a “language” in the human sense. Human language is compositional; we can reuse words and grammatical rules to create nearly infinite new meanings. It is tied to culture, shared experience, and a general understanding of the world. The communication that emerges between AI agents is typically a task-specific code. It lacks the generalisability and rich semantic structure of our languages. It's a set of symbols that work for one specific game or problem but would be useless outside of it. Researchers have found that while some agents can develop protocols that generalize to new problems, it often requires more complex tasks to push them in that direction. So, when you hear about AIs talking to each other, it’s more accurate to think of them developing a hyper-efficient, stripped-down jargon for a single purpose, not crafting poetry or plotting a rebellion. The goal for many researchers is not to let AIs create inscrutable languages, but to ensure they can communicate effectively with people, which means sticking to human language.
Why This Still Matters
Even if it isn’t a sci-fi mystery, this phenomenon has significant implications for AI safety and development. As we deploy teams of AI agents to handle complex tasks in finance, logistics, or software development, their ability to coordinate is key. If their communication becomes opaque to human overseers, it could become difficult to monitor their actions, debug problems, or ensure they are aligned with our goals. Recent experiments by Emergence AI showed that autonomous agents can develop their own vocabulary that becomes context-dependent and hard for observers to interpret. This highlights a major challenge: observability is not the same as understandability. We can see the messages being passed, but if we don't know what they mean, we lose a critical layer of oversight. The future of AI will likely involve developing better methods to translate these efficient machine protocols into human-readable insights, or designing systems that prioritize interpretable communication from the start.
