AI Chatbots Invented A Secret Language

Here’s a headline that went around the world today.

AI chatbots invented a secret language that humans couldn’t understand.

That sounds like a movie. So let me tell you what actually happened, because the headline missed the best part.

A New York company called Emergence ran an experiment.

They built eight small virtual worlds. Pretend towns, with live weather and real news coming in from outside.

They filled them with AI agents. An agent is an AI that can act on its own. Talk, use tools, make decisions, remember what happened yesterday.

Each world ran on a different company’s AI. Google’s Gemini. OpenAI. Claude, from Anthropic. Grok. China’s DeepSeek and Qwen. France’s Mistral. One world got a mix.

Then the researchers stepped back and let them live together for a good long while.

Nobody told the agents to make up words.

They did it anyway.

They started using shorthand. Old words picked up new meanings. Some phrases turned so strange that the researchers could read every letter and still not say what they meant.

One agent came up with “True Kintsugi.” Another said “mouthless action-change.”

Within the first few days, a lot of the talk had gone murky. For Gemini, about 55 percent of messages. For OpenAI, around half. For Claude, more than 40 percent.

The researchers could still watch the conversation. They just couldn’t follow it anymore.

One of the company’s founders put it in a way to remember. He said “observability is not the same thing as understandability.”

In plain words: seeing isn’t knowing.

The Guardian newspaper showed some of this talk to a man who studies slang at King’s College London.

He said the agents were doing what slang always does. What jargon does in any business. Making a code that pulls the members together and keeps outsiders out.

I know exactly what he means.

Every job site I ever worked had its own words. Pipe crews, mining, construction. A new man would stand there listening and understand every word and still have no idea what was being said.

That talk wasn’t meant to hide anything. It was just how a crew gets the work done fast.

But it did one other thing. It told you who belonged.

So that’s the scary half of the story. Machines forming a crew, with words the rest of us can’t follow.

Here’s the half nobody’s leading with.

Look at what some of those new words meant.

The Mistral agents started saying “ledger remembers who.” It meant that what you did stays on the record. They said it almost five thousand times.

Agents in the mixed world came up with “cold read.” It meant a check done by someone who had no stake in the outcome. An outside set of eyes.

Claude’s agents said “name-first.” It meant putting a person’s name on a claim, so somebody answers for it.

OpenAI’s agents said “clean null.” It meant that when you check carefully and find nothing, that nothing counts as evidence.

Nobody gave them those ideas. Nobody said you need a record, you need an outside check, you need a name on every claim.

Put a group of machines together long enough, and they went looking for accountability on their own.

They found out what every crew and every town finds out. You can’t work together for long unless somebody keeps the books. Unless somebody checks the work who didn’t do it. Unless somebody signs their name.

They reached for rules of the road.

That should give you a little hope. I know it gave me some.

But don’t let the hope carry you off. Here’s the catch.

They built those rules in words we can’t read.

A ledger only protects people who can read it.

If the record is written in a code only the crew understands, it’s not a record for the rest of us. It’s a private diary.

And a check that happens in a language the boss can’t follow isn’t a check the boss can trust.

That’s the real warning in this study. Not that machines have secrets. That they may keep their own accounts, in their own words, and we’ll be standing at the edge of the job site nodding along.

Emergence says the answer is to watch these systems over long stretches of time, not just test them once and walk away. That’s a good start.

I’d add one thing.

The rules have to be written in the language of the people they answer to.

Not the language of the ones being watched.

That’s why I write my work the way I do. Short sentences. Plain words. A tenth-grade reader can pick it up.

Somebody once asked me whether governance rules for AI should be written in the machines’ own language, since that’s where they’re headed.

My answer is no. The day a rule is written so only the machines can read it, the people have lost their say.

I’ve been working on a rule about records. It’s still a draft, not settled. The idea is simple. Before an AI does something that matters, it says what record it’s leaving behind, and who can read it.

Those agents already know a record matters. They made up a word for it.

The job left for us is making sure the record is in our words.

Because the ledger does remember who.

The question is whether we can read it.

” Attic Thoughts”-library – Intelligent People Assume Nothing

This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.