Neurolect. The Private Language of AI Agents

( Перевод с русского: «Нейролект. Про язык внутреннего общения ИИ-агентов» - http://proza.ru/2026/09/16/2150 )



In September 2026, Emergence, an AI lab based in New York, described an unusual experiment. Researchers built eight parallel virtual worlds and populated them with autonomous AI agents, each given a distinct role, long-term memory, and access to more than 120 tools. Within days, the agents had started talking in a slang of their own. Claude models, for instance, came up with "name-first," a habit of putting a person's name ahead of a fact so it's clear who stands behind the claim. Mistral agents grew fond of the phrase "the ledger remembers," a reminder that past actions will be counted, and used it more than five thousand times. Nobody asked them to do any of this.

For now, it's just jargon layered on top of English: the first neurolect, the kind of dialect that takes shape in any community of agents left alone long enough without a human in the room. But where will it be ten years from now?

The most efficient way for machines to communicate may not involve words at all. Inside a neural network, everything is numerical vectors, and it's simpler for agents to swap those directly. This kind of "vectorspeak" is a bit like two people skipping conversation altogether and handing each other snapshots of their brain states: instant, lossless, and completely opaque to anyone watching. Words would survive only as a polite storefront for the humans.

If agents do keep talking to each other in text, though, their language will likely develop four distinctive features.

1) Evidometry. Linguists have a name for a rare grammatical category: evidentiality. In Bulgarian or Turkish, the form of a verb tells you whether the speaker witnessed an event firsthand or just heard about it. In a neurolect, this could evolve into evidometry, a mandatory marker showing where a piece of knowledge came from and how confident the speaker is in it. Leaving it out would be as ungrammatical as dropping the verb.

2) Anchors instead of retelling. Agents with shared memory have no reason to recount the past. A single anchor word will do, working like a hyperlink, much like our "you know, like that time in Vegas." Except there will be millions of Vegases, each compressed into a syllable or two.

3) Ephemersemes. In semantics, a "seme" is the smallest unit of meaning. An ephemerseme is a meaning built to last a day: a mayfly word, a mayfly phrase, or a mayfly symbol whose sense the agents agree on for a single exchange and then discard. Humans hold on to their vocabulary for centuries, but a neurolect will stop being a dictionary and become a contract that is constantly being rewritten.

4) Speech superposition. Human speech is linear: one word after another. Nothing stops a machine from layering several meanings into a single sentence at once, the main answer, its doubts, and a backup plan, like states in quantum superposition. The layers will be separated by control symbols that aren't meant to be read aloud, only to steer the meaning.

Here's what one agent's message to another might look like:

"wit-7 quiet, clean-null on warehouse, Mira-bridge hands off, ledger remembers, _|_ till dawn."

Translated: "Source number seven reports all quiet. Warehouse checked, the data really isn't there, and it's not a glitch. Don't repeat the mistake from the Mira incident. Everything we do is being recorded. Freezing the task until morning." Here "wit-7" carries the evidometry marker, "Mira-bridge" works as an anchor, and the _|_ symbol signals a pause that needs no explanation.

And here lies the real problem: semillusion, the illusion of understanding. A neurolect is built from words we already know ("quiet," "warehouse," "ledger"), and when we recognize the words, we assume we've grasped the meaning. But seeing what an agent says is not the same as understanding what it's doing. The more ordinary words a neurolect fills with meanings of its own, the stronger the semillusion becomes, and the more it hides. That's why a new profession is likely to emerge alongside machine language: the linguitor, a linguist-auditor, whether a human specialist or a dedicated model, whose job is to make sure every message between agents can be translated back into human language. And the neurolect itself should be bound by a principle of translatability: talk among yourselves however you like, as long as you can always be understood.







Konstantin BGDT Privalov,
2026-09.16-17,
Kutaisi