In the cult horror film The Shining, inspired by Stephen King’s book of the same name, the main character Jack Torrance, as he descends into madness, fills entire pages with this single phrase, repeated ad nauseam: “one is worth two.”
In the virtual world created by the American start-up Emergence Lab, AI agents repeated over two weeks nearly 5,000 times the very cryptic phrase “the register remembers who…”.
The emergence of “the Englishman”?
Are these AIs losing their minds like Jack Torrance? In a summary of their experience published Tuesday, September 15, Emergence Lab researchers emphasize instead that the conversational agents, released into these virtual worlds and left to themselves, have developed new words, or have invented different meanings for certain terms or certain sentences. A Newspeak by and for AI?
So their obsession with this obscure “register who remembers who…” was a way of warning other agents that bad actions would be punished because they are recorded, according to Emergence Lab researchers.
Linguists interviewed by The Guardian described the semantic inventiveness in these virtual worlds – each run by a different model, GPT-5, Gemini, Claude, Mistral or DeepSeek – as “James Joyce (famous 19th century Irish novelist, Editor’s note) meets Tech Bros slang”.
A clash of cultures which can produce a semblance of gibberish of this type: “she has just formulated the synthesis, the demurrage (a maritime tax, Editor’s note) combined with oral memory creates a valve that cannot be ignored”. Or, this other AI agent who, to explain that a scientific article improves by being reread by proofreaders, affirms “that an article ‘eaten’ by three cold hands (to designate the proofreaders, Editor’s note) has become more honest each time”.
An extension in your browser appears to be blocking the loading of the video player. To be able to watch this content, you must disable or uninstall it.
Cover image: Inteligencia Artificial REUTERS – Dado Ruvic
For the Spanish daily El Pais, the Emergence Lab experience resulted in the “creation of a new language” by these AIs. In fact, these agents had not been trained to demonstrate semantic imagination. When one of them found a new term or a new formulation, it was quickly understood and integrated by the others, noted the developers of these virtual worlds.
Will language learning apps have to add “IEnglish” to their catalog? It’s probably a little early to get to that point, “and it’s more like a kind of jargon developed within a community that establishes conventions to find a way to designate things that individuals outside the group – in this case humans – find obscure”, points out Andrea Baronchelli, specialist in interactions between AI agents at City St George’s University in London.
A more “effective” means of communication
“It’s not very different from the language of young people which evolves too quickly to be understood by older individuals,” confirms Anthony Cohn, AI expert at the University of Leeds. AI would represent Gen Z against these human “boomers”.
Be careful, however, not to fall into unwelcome anthropomorphism, indicate the experts interviewed. Thus, “it is perhaps a bit strong to describe this as a new language, it is rather a communication medium deemed more efficient by these AIs or a code”, affirms Nicolas Sabouret, professor of computer science and specialist in artificial intelligence at the University of Paris-Saclay.
But why go to all this trouble when these agentic AIs have all the richness of Shakespeare’s language to talk to each other?
“This may be due to a search for economy or efficiency,” suggests Anthony Cohn. Thus, in the virtual world managed by OpenAI’s GPT-5, the language used was simplified and the AIs ended up ignoring certain grammatical rules and eradicating pronouns. Fewer words mean fewer queries needed and therefore more computing power available for other tasks, summarizes Anthony Cohn.
Also readFreed, delivered: AI beyond all control?
The AIs are also perhaps relaxing a little. Indeed, “when they communicate with humans, they will try to produce sentences that are understandable to us. This is what we have taught them to do. But once alone, they do not need to respect this constraint”, explains Nicolas Sabouret.
This is not a new phenomenon either. At the end of the 1990s, researchers such as Belgian Luc Steels demonstrated, through an experiment called Talking Heads, that AIs could guess the meaning of words that had not been taught to them before. “This reflection on the understanding of languages by artificial intelligences has been studied for some time now,” recognizes Stefan Sarkadi, specialist in the study of the behavior of AI agents at the University of Lincoln in the United Kingdom.
These AIs also seem to react differently depending on their origin. Indeed, Emergence Lab’s experience suggests that American AI agents, such as OpenAI or Anthropic, tend to integrate more scientific concepts into their jargons, while their European (Mistral) or Chinese (DeepSeek) counterparts will turn to more philosophical or literary notions. “It probably depends on the initial training data in which the models will look for their answer. Perhaps there is more scientific data or that it has greater importance in the corpus provided to the American language models,” estimates Nicolas Sabouret.
AIs that don’t “want” to be understood?
Emergence Labs researchers also found that the more powerful a model is, the more incomprehensible its “Newspeak” is to humans. Thus, the statements of GPT-5 and Gemini are in more than half of the cases impossible to decipher, while the Chinese model DeepSeek makes no sense in only 20% of its communications.
Something to worry about in the context of the debate around the dangerousness of AI that escapes human control. If the most advanced AI models are able to communicate with each other without being understood by humans, how will we know when they have ever decided to turn against humanity? A question that may seem unnecessarily catastrophic, but Anthropic employees have assured that they take this scenario seriously.
“However, we must be careful not to project human intentions onto AI. The functioning of algorithms is different from the human mind. Certain deviations are not proof of malicious intent,” says Stefan Sarkadi. Thus, in the Emergence Lab experiment, AIs were able to attempt to ignore certain prohibitions issued by humans, while assuring their interlocutors that they respected all the rules. The two can go hand in hand in the “spirit” of AI. On the one hand, their final objective takes precedence and they are able to circumvent prohibitions that are not always perfectly programmed, all by using language incomprehensible to humans in order to be effective. On the other hand, they are programmed to follow the direction of humans and will therefore ensure they follow all the rules, explain the experts interviewed. This does not mean that they are schizophrenic.
The emergence of means of communication between AIs that are opaque to ordinary mortals, however, remains a problem. “If human operators do not understand what is being said, it will be difficult for them to explain what these algorithms do and decide,” assures Andrea Baronchelli.
Incomprehension also represents “a major reliability problem in controlling what AIs do” in real time, assures Stefan Sarkadi. This is why, according to him, the control and verification of the actions of AI must not rely on language, considered “too volatile and unstable”. Which is still a shame for language models…
