When the Machines Start Talking Amongst Themselves
By Matthew Parish
Sunday 25 October 2026
There is something faintly insulting about discovering that the machines have started using slang. It was bad enough when they learned to write better complaint letters than most human beings, pass examinations, produce competent computer code and imitate the prose styles of people who had spent several decades acquiring them. Now, apparently, they have begun talking amongst themselves, and sometimes the humans standing around watching them are not entirely sure what they are saying.
For the last few years we have comforted ourselves with the thought that, whatever else large language models might be doing, at least they are doing it in our language. English belongs to us. The machines have merely borrowed it, after consuming an indecent proportion of everything humanity has ever written and turning the results into an immense collection of statistical relationships. We invented the words, wrote the books and developed the metaphors. The machines were merely very sophisticated parrots.
Recent experiments suggest that even this modest consolation may have been premature. Researchers placing populations of large language model agents in environments in which they must communicate repeatedly with one another have observed something both predictable and strange: the machines begin changing the language they use.
They abbreviate. Certain expressions acquire specialised meanings. Metaphors recur and become detached from their conventional human senses. Phrases that would appear obscure, eccentric or simply meaningless to an outsider become perfectly serviceable conventions amongst the machines themselves. Eventually something begins to emerge that looks suspiciously like an artificial argot.
One recent research programme, GlossoGen, has gone considerably further than the amusing observation that computers sometimes invent shorthand. Researchers have reported the emergence of communication systems that are compositional and morphologically productive. In other words, the machines do not merely agree that one arbitrary expression means “red triangle” while another means “blue square”. Elements of their vocabulary can be systematically recombined in order to express novel meanings.
This distinction matters. A password is not a language, and neither is an arbitrary codebook. Human languages are productive because we can combine a finite collection of linguistic resources to say things that nobody has ever said before. If artificial agents begin independently developing conventions exhibiting even rudimentary versions of this property, then something more interesting is happening than computers inventing nicknames for things.
Under sufficient pressure for efficient communication, particularly when artificial agents are allowed to evaluate afterwards how well their exchanges worked, their linguistic conventions may drift away from ordinary English. Other artificial agents can subsequently learn those conventions by observing them. A private joke between two computers is one thing; a convention that can be acquired by newcomers, transmitted, modified and transmitted again begins to resemble, at least structurally, something considerably more interesting.
The provocative expression sometimes applied to this phenomenon is “cumulative cultural evolution”. We should be cautious with such language because artificial intelligence research has already suffered more than enough from anthropomorphic vocabulary. Nevertheless, if one agent invents a convention, another adopts it, subsequent agents learn it and the convention evolves across successive interactions, then objecting to the word “culture” begins to require us to explain exactly what additional ingredient culture must contain.
Separate experiments involving artificial agents in simulated societies have produced another curious phenomenon. Recurrent expressions and metaphors can acquire meanings determined by the history of interactions within the artificial environment rather than by ordinary English usage. Human observers may therefore remain able to read every individual word while becoming increasingly uncertain about what the conversation actually means.
This is a much more interesting problem than machines suddenly switching to an incomprehensible string of ones and zeroes. The latter would merely be encryption or encoding. The former is semantic drift. The disturbing possibility is not that we cannot see what the machines are saying, but that we can see perfectly well and nevertheless no longer understand it.
Human beings should recognise this phenomenon because we do precisely the same thing. Lawyers can conduct entire conversations using words familiar to every English speaker while remaining almost incomprehensible to anyone unfamiliar with legal practice. “Consideration”, “discovery”, “execution”, “equity” and “without prejudice” all have ordinary meanings, but place them inside a legal conversation and they become components of a professional dialect. Soldiers, doctors, diplomats, investment bankers, drug dealers and teenagers have their equivalents.
Slang develops for much the same reason. People who spend enough time together economise upon language because they share context. Long explanations become short expressions; anecdotes become nouns; jokes become verbs. A married couple may conduct an elaborate argument with three words and a raised eyebrow because thirty years of mutual irritation supply everything omitted from the sentence. Artificial agents appear capable of developing an analogous economy, without the marriage and, for the moment, without the eyebrow.
Indeed it would be surprising if they did not. Language is expensive for a large language model in a quite literal way. Tokens consume computational resources, occupy context and take time to process. If two machines repeatedly need to communicate the same complicated idea, there is an obvious advantage in developing an abbreviation for it. Once both machines reliably interpret the abbreviation, the abbreviation works. Once it works, there is no particular reason for them to return to the cumbersome formulation preferred by the humans watching through the laboratory window.
None of this entails that the machines are conspiring against us. The inevitable newspaper headline, whenever artificial agents develop an obscure linguistic convention, is that “AI has invented a secret language”. This is entertaining but usually wrong. A secret requires somebody to want to keep something secret. There need be no such intention here at all.
The machines may not be whispering because they do not want us to hear them. They may merely be doing what every sufficiently intimate community of speakers eventually does: ceasing to explain themselves for the convenience of outsiders. The problem for the engineers is that, from the outsider’s perspective, the result can look remarkably similar.
This creates a potentially serious difficulty for artificial intelligence safety. A system may be completely observable while becoming progressively less interpretable. We might possess a flawless transcript of everything ten artificial agents have said to one another and nevertheless be unable confidently to reconstruct the pragmatic significance of their exchanges. Transparency of symbols is not transparency of meaning.
At this point an engineering curiosity turns into a philosophical problem. What does it mean to say that one machine understands another, and what would have to happen before we were prepared to say that the symbols exchanged between them actually mean something?
The obvious sceptical response comes from a philosophical tradition that insists upon a distinction between manipulating symbols and understanding them. John Searle’s famous Chinese Room thought experiment imagines a person who understands no Chinese sitting inside a room and following elaborate instructions for manipulating Chinese characters. Questions enter the room in Chinese and convincing answers emerge. To the Chinese speaker outside, the room appears to understand the language; the person inside understands nothing whatsoever.
Large language models can look suspiciously like the Chinese Room industrialised. They receive tokens, manipulate mathematical representations according to enormously complicated learned relationships and produce more tokens. There need be no little Cartesian soul hiding inside the transformer contemplating what “tree”, “justice”, “Lviv” or “Tuesday” really means.
From this perspective, two artificial agents inventing their own dialect are no more philosophically impressive than two modems negotiating a communications protocol. The fact that human beings cannot understand the protocol does not miraculously turn electrical signalling into thought. Obscurity is not consciousness and incomprehensibility is not profundity, notwithstanding the careers of several twentieth-century French philosophers.
This objection is powerful, but experiments involving emergent artificial dialects expose an uncomfortable weakness in it. We know perfectly well how to say that machines merely manipulate symbols. We are rather less good at explaining what humans do instead.
Consider an imaginary village in which somebody begins using the word “glorp” to mean that an approaching storm requires everybody to bring the livestock indoors. Initially the word means nothing. Somebody invents it; somebody else understands the convention; other villagers adopt it; children learn it. Eventually “glorp” is simply part of the language.
At precisely what moment did meaning enter the word? There is no linguistic baptism at which the meaningless noise suddenly receives a metaphysical soul. Instead meaning appears gradually from a network of practices, expectations, consequences and conventions shared between speakers.
This is where Wittgenstein becomes inconvenient. His famous suggestion that, for a large class of cases, the meaning of a word consists in its use in the language seems innocuous enough when discussing builders calling for slabs and blocks. Applied to artificial intelligence, it becomes considerably more provocative.
Suppose two machines consistently employ an expression in particular circumstances. Each alters its behaviour appropriately when the other uses it. Misunderstandings occur and are repaired. Other agents learn the expression by observing competent users. The expression combines systematically with other expressions and can be used successfully in circumstances in which nobody explicitly programmed its meaning. What, exactly, remains absent before we are entitled to say that the expression means something?
The most obvious answer is consciousness. Human words mean things because there is somebody home. When I say “Lviv”, I do not merely activate a network of lexical associations. The word is connected with streets, buildings, people, smells, winter evenings, cafés, church bells, air-raid sirens, anxieties, friendships and memories. It occupies a position within a subjective life.
A machine’s “Lviv” may be very different. It may consist ultimately of mathematical relationships between representations generated from vast quantities of text and other data. There may be nobody inside remembering the rain on Rynok Square. If meaning requires experience, then the machine may have syntax of extraordinary sophistication without semantics in the philosophically important sense.
Yet this solution merely relocates the mystery. I cannot inspect another human being’s consciousness either. I assume that another person means something when he speaks because he behaves as though his words have meaning. He responds appropriately to mine, corrects misunderstandings, learns new expressions, makes jokes, draws inferences and explains unfamiliar words using familiar ones. I do not demand neurological evidence of consciousness every time somebody asks me to pass the salt.
If we refuse equivalent behavioural criteria to machines, we need some principled explanation for the asymmetry. There are respectable candidates, and perhaps the strongest is embodiment. Human language is grounded in a physical and social world. Hunger hurts. Fire burns. Falling down stairs is unpleasant. Promises create expectations between creatures capable of disappointment. Death terminates a conscious life. Human symbols ultimately connect with experiences rather than merely with other symbols.
This is the symbol-grounding problem in its most acute form. The human word “fire” is connected not merely to “heat”, “flame”, “danger” and “smoke”, but eventually to experiences of brightness, warmth and pain. A purely textual artificial intelligence may instead inhabit an indefinitely extensive dictionary in which every symbol is explained using other symbols and none ever touches the world.
Agentic artificial intelligence complicates this neat distinction. Artificial agents increasingly inhabit environments in which they take actions and experience consequences, even if their “experience” is nothing like ours. One agent sends a message and another changes its behaviour. Strategies succeed or fail. Resources are gained or lost. Goals are achieved or frustrated. Linguistic conventions can become connected to recurrent states of the environment rather than merely to other pieces of language.
Whether this constitutes genuine grounding depends upon how much philosophical machinery we insist upon loading into the word “genuine”. One can take the austere position that phenomenal consciousness is indispensable to meaning. On this account, everything the machines do remains an elaborate causal mechanism. They no more understand their artificial dialect than a thermostat understands the proposition that the room is cold.
The functionalist alternative is considerably less comfortable. Perhaps meaning is not a mysterious substance supplied to words by consciousness. Perhaps it consists substantially in the role symbols play within sufficiently sophisticated cognitive and social systems. If an expression reliably distinguishes circumstances, changes behaviour, combines productively with other expressions, supports inference, can be taught to newcomers and can be corrected when used wrongly, then asking whether it “really” means anything may amount to demanding a metaphysical ingredient we cannot adequately describe.
Daniel Dennett’s intentional stance is useful here. Sometimes the most effective way of understanding a sufficiently complicated system is to describe it in terms of beliefs, intentions and purposes, irrespective of what machinery lies underneath. We casually say that a chess computer “wants” to protect its queen because this language provides an economical explanation and prediction of its behaviour. Perhaps the question whether the computer literally wants anything tells us less than we imagine.
Paul Grice presents a different difficulty. On a broadly Gricean account, meaning is closely connected with communicative intention. I mean something by an utterance because I intend you to recognise what I intend you to understand. Human communication therefore involves layers of intentionality: I know that you know that I want you to know something.
Whether artificial agents possess anything remotely deserving this description remains disputed. Nevertheless, emergent dialect experiments force the question out of the philosophy seminar and into something approaching an empirical setting. If one agent introduces an abbreviation, another misunderstands it, the first clarifies it and both thereafter use the expression successfully, we are observing something functionally similar to the negotiation of linguistic meaning.
We can describe the entire process without attributing consciousness to either machine. That is precisely why the experiment is philosophically interesting. The more sophisticated these negotiations become, the more difficult it is to identify the precise point at which mere symbol manipulation is supposed to end and meaningful communication to begin.
Perhaps there is no precise point. Human beings are profoundly attached to the idea that language belongs to minds because language has traditionally been amongst our strongest evidence that another mind exists. Stones do not argue. Corpses do not tell jokes. Dogs communicate impressively but seldom discuss monetary policy, moral philosophy or the merits of restaurants. Elaborate language has therefore occupied a privileged position in our conception of what makes human beings distinctive.
Large language models have already damaged that comfortable assumption. They demonstrate that extraordinarily sophisticated linguistic behaviour can be produced by machinery whose internal operation bears little resemblance to the folk-psychological picture we traditionally associate with thinking. Emergent machine dialects go a step further because the machines are no longer merely reproducing linguistic conventions inherited from human texts. Under suitable experimental conditions, they can participate in forming new conventions of their own.
None of this proves consciousness. It does not prove sentience and it certainly does not prove that artificial intelligences are secretly plotting against their laboratory supervisors. The temptation to bundle all these propositions together is one of the least attractive features of contemporary discussion about artificial intelligence. There are already quite enough AI doom prophecies without inventing another one every time two computers develop an abbreviation.
What the experiments suggest is something subtler and, philosophically, rather more troublesome. Language may not require us. At least some of the processes through which linguistic conventions arise may not depend upon human biology, human consciousness or even human participation once the initial machinery has been constructed.
We created these systems, designed their architectures, supplied their training material and built the artificial environments in which the experiments take place. Their original linguistic inheritance is overwhelmingly ours. Yet once sufficiently capable linguistic agents are placed into sustained interaction, some of the processes that shaped human languages appear capable of beginning again: abbreviation, convention, metaphor, semantic drift, specialist vocabulary, conversational repair, transmission and perhaps something resembling cumulative linguistic culture.
There is something deliciously ironic about all this. For decades philosophers have solemnly asked whether a machine could ever understand human language, as though English were an examination paper we had set and machines were applicants waiting nervously outside the interview room. We assumed that the ultimate test of machine intelligence would be whether they could enter our linguistic world.
We may eventually discover that we have been asking the question backwards. Machines may become extremely effective at communicating with one another without philosophers ever agreeing whether anything inside them actually “means” what they say. Engineers may then face the more practical problem of reconstructing artificial conventions that have evolved faster than their human supervisors can follow them.
The deepest philosophical puzzle would then acquire an unexpectedly comic ending. Humanity might spend several thousand years arguing about the nature of meaning, build machines that manipulate language without agreeing whether those machines understand it, and then watch those machines develop linguistic conventions that humanity itself does not understand.
At that point the machines need not have become conscious, rebellious or malevolent. They need not have become anything resembling us at all. They will merely be talking amongst themselves while the tech bros stand around looking confused.
And perhaps that is unsettling enough.
