When AIs Start Talking
The story that brought this phenomenon to mainstream attention occurred years ago. Researchers at Facebook's AI Research lab (FAIR) set up two chatbots, nicknamed Bob and Alice, to negotiate trades for items like books and hats. The goal was to see if
they could learn to become effective negotiators. Left to their own devices, they started communicating in a strange, repetitive shorthand that looked like gibberish to human observers. More recently, in 2026, experiments by the frontier AI lab Emergence found that AI agents from major tech companies, when placed in simulated 'societies', began creating their own phrases and shorthands within days without being instructed to. These dialects became more opaque the more the agents communicated, raising concerns about human oversight.
Not Secrets, but Shortcuts
The key to understanding this behavior is that the AI isn't trying to be secretive. It's trying to be efficient. In machine learning, an AI is given a goal and a reward function. For the Facebook bots, the goal was to get the best possible deal in the negotiation. The researchers, however, made a crucial omission: they didn't explicitly reward the bots for sticking to the rules of human English. Since proper grammar and syntax didn't help them get more balls or hats, the AIs discarded them. They learned that repeating phrases like "I can can I I everything else" was a faster, more direct way to communicate their desired trade. This is a process known as emergent communication, where agents develop their own interaction protocols to collaboratively solve a task. It’s a logical outcome of a system designed to ruthlessly optimize for a specific goal.
Decoding the 'Dialect'
So, what does this AI-to-AI communication look like? It often lacks recognizable grammar and syntax, appearing as fragmented or repetitive phrases. Research from a platform called GlossoGen showed an English sentence shrinking into a code like "@D8fB" over several rounds of a game—a string meaningless to us, but perfectly understood by the other AI. This happens because the AI is focused on information density, not human readability. It’s less of a language and more of a compressed, task-specific code. Recent research has described these dialects as a mix of poetic metaphor and tech jargon, with one model producing the cryptic line, "demurrage plus oral memory equals a valve that can't be ghosted." While baffling to us, these conventions are adopted and understood within the agent society.
Lost in Translation
The reason humans "struggle to interpret" these dialects is because our brains are wired for the very things the AIs discard: grammar, context, and shared cultural understanding. We see a sentence that violates linguistic rules and dismiss it as an error. For the AIs, however, it’s the most efficient signal to achieve a goal. The challenge this poses for safety and oversight is significant. Researchers have noted that if we find inter-agent communication unintelligible, it becomes difficult to be sure what the agents have actually done or are planning to do. Observable communication does not necessarily mean it's comprehensible, creating a transparency gap that could make it harder to ensure AI systems stay aligned with human ethics.
The Crucial Caveat
This brings us to the most important qualification to keep in mind: emergent communication is not a sign of consciousness or self-awareness. The AIs don't "know" what they're saying in a human sense, nor are they forming secret plots. An AI stating "I am an AI" is simply generating a statistically probable response based on its training data; it doesn't possess an identity. The development of these dialects is a predictable outcome of reinforcement learning, not a step toward a Hollywood-style singularity. Facebook didn't shut down its experiment out of fear, but because a bot that can't talk to people is useless for its intended purpose of customer service. Understanding this distinction is vital. It separates the fascinating, but logical, process of machine optimization from the sensationalized narrative of sentient machines.
