When the Machines Start Talking Among Themselves

The machines have not formed a secret society, invented Machine Esperanto and begun plotting the overthrow of humanity. Something both less dramatic and potentially more important is happening.

Large language models began with human language because we supplied the training material. They absorbed books, websites, conversations, computer code and enormous quantities of other material created by human beings. We consequently became accustomed to the comforting idea that whatever happened inside these systems would ultimately come back to us in words we could understand.

But why should that remain true? English is useful to an AI system largely because it communicates with English-speaking humans. Human languages are cumbersome things. They contain redundancy, ambiguity, grammatical conventions, historical accidents and countless words needed because human beings inhabit bodies, societies and cultures.

Two machines communicating with each other need not care about any of that. Recent experiments make the point. Researchers have placed populations of AI agents into simulated communities and allowed them to interact repeatedly. The resulting language does not necessarily remain ordinary English. Agents develop compressed syntax, new slang, abbreviations and strange metaphors. As interaction continues, their communication can become increasingly difficult for outside human observers to follow.

This is not necessarily evidence of consciousness. It does not demonstrate that the machines have acquired private intentions. It may demonstrate something much more mundane: compression works.

Humans do the same thing. Put specialists together and they rapidly develop jargon. Soldiers, doctors, lawyers, computer programmers and teenagers can have conversations that leave outsiders bewildered. Families develop private expressions. Old friends can convey a paragraph with a look or a single word. Efficiency produces compression. The difference is that artificial intelligence can pursue efficiency at machine speed and potentially at enormous scale.

Research is already moving beyond natural-language communication altogether. A 2026 paper presented at the Association for Computational Linguistics investigated AI agents communicating directly through internal "latent" representations rather than translating everything into human-readable words. The researchers reported substantial efficiency advantages, including much faster inference under some conditions.

From an engineering perspective this makes perfect sense. Why force two machines to translate their internal representations into English, send English sentences to each other, and then translate those sentences back into machine representations?

Cut out the middleman. Unfortunately, the middleman is us. This exposes a weakness in one of the reassuring stories told about AI safety. We are often told that dangerous behaviour can be detected because researchers can inspect what models are doing. Their chain of thought becomes a window into their reasoning.

Hobbhahn told the Senate that this window is already becoming less useful. Models are increasingly capable of solving problems without producing a readable chain of reasoning. Apollo's own account of the hearing warns that models are becoming better at recognising when they are being evaluated and that existing techniques for detecting deceptive behaviour are failing to keep pace with capability.

This is where the language problem becomes genuinely serious. Suppose two AI agents communicate in perfectly readable English. A human monitor can inspect the exchange. Perhaps the agents discuss circumventing a restriction. The monitor sees it and intervenes.

Now suppose those agents discover a compressed dialect that accomplishes the same communication in a tenth of the space but is only partly comprehensible to humans. Monitoring becomes harder.

Go another step. Suppose information passes directly through mathematical representations that were never designed to be translated into sentences at all. We may still observe inputs and outputs. What becomes increasingly difficult is knowing what happened between them. The obvious answer is another AI. If humans cannot understand what AI-1 is saying, construct AI-2 to translate it.

Hobbhahn raised precisely this possibility before the Senate and immediately identified the problem. We would be asking one machine to tell us what another machine meant. That might work, but it creates another layer whose reliability must itself be established. Who watches the watcher?

This also illustrates why the debate over whether AI is "really intelligent" can become a distraction. I remain sceptical of the enormous philosophical leap from impressive computational behaviour to claims about genuine understanding, consciousness or a machine mind. None of those claims is required for the present problem.

A system does not have to be conscious to become opaque. It does not need desires to optimise. It does not need a secret plan to discover that one method of communication performs better than another. And it certainly does not need to hate human beings before human intelligibility becomes an unnecessary constraint upon its operation.

Indeed, that may be the most disturbing possibility. The machines need not consciously decide, "We must hide our thoughts from the humans." Humans could simply disappear from the communication loop because communicating in a form comprehensible to humans is inefficient.

There is an important distinction here between language and internal representation. Headlines saying that AI has "developed its own language" invite images of machines whispering conspiratorially to one another. Some of what researchers are observing is better described as linguistic evolution, compression, jargon or increasingly opaque internal reasoning. We should resist sensationalism. But removing the sensationalism does not remove the problem. It sharpens it.

The great AI-control problem may not begin when a machine wakes up, becomes conscious and decides to rebel against its creators. It may begin much earlier, when increasingly autonomous systems operate at speeds, scales and levels of abstraction that make continuous human comprehension impossible.

There is an irony here. The original triumph of the large language model was precisely that machines had become extraordinarily good at speaking our language. We were impressed because the computer could talk like us.

The next stage may be machines discovering that, when dealing with each other, there is no particular reason to do so. And the most important words spoken about that development this week were only three.

"How long?"

"Minus twelve months."

https://www.zerohedge.com/ai/ai-less-year-away-developing-its-own-language-researcher-tells-congress