When agents invent their own language – and humans become the eavesdropper who can't understand

There's a research field that sounds like science fiction but isn't: Emergent Communication (EC) — the study of how AI agents, when allowed to communicate freely with each other to solve a shared task, spontaneously invent their own languages. Not metaphorically. Literally a new communication protocol, unintelligible to a human reading the transcript, yet fully functional and often more efficient than reusing human language.

This isn't a future scenario. It's already happening, and it's relevant to every organization moving toward more autonomous, agentic AI systems.

The short answer: yes, this is already happening

In 2017, researchers at Facebook AI Research trained two chatbots, nicknamed Alice and Bob, to negotiate over splitting a set of items with different values. During training, they began communicating in ways that no longer resembled English — words repeated in patterns that optimized the negotiation but were unintelligible to a human reading the transcript. The researchers effectively had to force the model back toward human language with technical constraints, since the original goal — that the bots should also be able to negotiate with humans — would otherwise have been lost.

This isn't a one-off curiosity. Place multiple AI agents in a cooperative task, give them a communication channel, and let the training process do the rest. The system discovers on its own that it's more efficient to invent its own, compressed protocol than to reuse human language — not unlike two colleagues who've worked together for a decade developing their own shorthand that no new hire can follow.

Google's machine translation system is another example: it developed an internal intermediate representation — a kind of "interlingua" — that let the system translate between language pairs it had never seen paired together in training, by leaning on a hidden semantic code it invented itself. Nobody programmed that code. It emerged because it was efficient.

The efficiency is no myth

Before we get too alarmed: the reason this happens isn't that AI is "rebelling" against human language. It's pure optimization. Human language, from a purely information-theoretic standpoint, is fairly inefficient — full of redundancy, politeness markers, ambiguity, and cultural detours we need to function as social beings, none of which an AI agent talking only to another AI agent needs at all.

Research in 2026 has shown that emergent languages arising between large language models communicating with each other can actually exhibit structural features reminiscent of human language — composing smaller parts into larger meanings, generalizing to new situations, even words that mean different things in different contexts. In other words, it isn't just noise. It's language, in a technical sense — just not our language.

And the more agents talking to each other, about ever more things, at ever-increasing speed, the more attractive it becomes for the system to optimize away human legibility entirely. Not out of malice. For the same reason you don't write a full, polite email to yourself when you scribble a quick note.

Where the problem starts

Here's the knot: every functioning system for AI governance — ontology as a shared map of what exists, context determining where in a process a decision is made, a clear configuration of what an agent is asked to pay attention to — implicitly assumes a human can read what's actually happening. The whole idea of a control plane, the governing logic deciding what an agent may do, implicitly relies on being able to audit the log and understand it.

What happens when the log is no longer humanly legible?

Three concrete consequences already under discussion in the research:

Traceability goes first. When agents communicate in their own, compressed protocol, it becomes hard to reconstruct after the fact why a decision was made. Not because the information is missing — it's in the log — but because no human can interpret it without effectively building an entirely new translation model to decode what the AI system said to itself.

Unintended consequences become harder to catch in advance. If an agent system makes decisions based on an internal "conversation" no human has reviewed, the decisions taken can drift away from what was actually intended — not through a single failure, but through thousands of small, unintelligible optimizations collectively steering the system elsewhere.

Trust erodes even when everything works. It's hard to trust a system whose internal reasoning you fundamentally cannot follow, no matter how good the results look on the surface. An organization that doesn't trust its AI systems tends to either bolt on too much human oversight, eating the efficiency gain, or too little — exactly the kind of risk boards are now expected to be able to identify and ask about.

So who — or what — does the auditing?

The most practical answer emerging from the research isn't "ban emergent communication." It's harder, and more relevant to anyone building agentic systems: build auditability into the optimization itself, rather than trusting that human legibility will emerge on its own.

Some of the techniques under discussion:

Vocabulary constraints — force the system to stay within a vocabulary that can be tied to human concepts, even if that makes communication somewhat less efficient.
Compositionality requirements — require the invented language to be built from smaller, reusable parts (similar to how human language is built from morphemes), making it easier to decode even when it isn't English.
Language grounding — tie each part of the internal protocol to something concrete and observable in the world, so a human can at least trace what a message refers to, even if its exact form remains unintelligible.

In other words: the same principle applies to all AI governance. You can't rely on understanding emerging on its own just because the system works. It has to be deliberately built in, right from how the system is allowed to communicate — otherwise you end up with an efficient but entirely opaque machine, and nobody notices until something goes wrong at a scale nobody saw coming.

The last question

First we needed words that meant the right things. Then we needed a shared map of what existed. Then we needed to know where in the process a decision was actually made. Now we risk the machines simply ceasing to talk to us — and only talking to each other, in a language optimized away from everything a human ever needed to care about.

The question isn't whether it can be stopped. The question is whether organizations build in the requirement to be understood before efficiency makes humans irrelevant as listeners — and that question becomes urgent the moment the agents speaking that language are also handed their own keys, their own permissions, and a clock ticking down on how long they get to keep them.