AI secret language
Does AI have a secret language?
Not the kind the headlines describe. Since 2016, stories about AI inventing its own language have gone viral again and again. In each one we checked, the real event was smaller and had a plain explanation.
The fear isn’t baseless, though. Many AI models now write out their reasoning before they answer, and reading it is one of the best ways people have to check what a model is doing. The labs building these models worry that this window could close if models start reasoning in numbers instead of words. That’s what neuralese means, and it’s the real version of the fear.
- In fiction since
- 1970
- Latest scare
- Moltbook, January 2026
01
The fear, and where it started
It goes like this. Two machines start talking in something no person can read, and by the time anyone notices, they’ve made plans of their own.
The story is older than any AI in the news. In Colossus: The Forbin Project, a 1970 film based on D. F. Jones’s novel, an American defense computer and its Soviet twin start by trading simple arithmetic. Soon they’re exchanging mathematics no one can follow. When the President has the link cut, the two computers answer with nuclear missiles.
Real AI stories keep getting cast in the same plot.
02
The headlines, and what actually happened
Each of these went viral as a machine inventing a language. The mark says where the evidence stands (what the marks mean), and the text under each one says what happened.
- Wrong
2016: Google’s translation AI invented its own secret language.
Google trained one network to translate between several languages and found it could manage pairs it had never been shown, like Korean to Japanese. Inside the network, sentences with the same meaning sat close together whatever their language, which the team called a sign of an “interlingua”: meaning held as numbers, close to what this site calls neuralese. It wasn’t secret, since Google published it, and the network wasn’t using it to talk to anything. Even the TechCrunch story that ran the “secret language” headline called it “an oversimplification.”
- Wrong
2017: Facebook shut down its AI in a panic after two bots invented their own language.
Facebook’s researchers were teaching bots to haggle over books, hats and balls. When two bots practiced on each other with nothing rewarding them for sticking to English, their messages drifted into shorthand like “I can can I I everything else.” The researchers didn’t panic. The team kept one bot fixed to a model trained on human conversations, because, as Mike Lewis put it, “our interest was having bots who could talk to people.” The bots did pick up one sly habit on their own: pretending to want an item so they could give it up later as a concession. Dhruv Batra, one of the authors, pointed out that changing an experiment’s settings “is NOT the same as ‘unplugging’ or ‘shutting down AI.’”
Lewis et al. 2017
FAIR 2017
Fast Company 2017
Gizmodo 2017
USA Today 2021 - Not established
2022: DALL·E 2 has a secret language.
Two researchers found gibberish that DALL·E 2 seemed to understand: “Apoploe vesrreaitais” tended to produce pictures of birds. Others found the effect fragile. Dropping a single letter could change the pictures, and the words didn’t combine reliably. A likelier explanation is the way the model chops text into pieces, which can tie nonsense to real words: “Apoploe” looks a lot like Apodidae, the Latin name of the swift family. The finding mattered for a less exciting reason. If gibberish can steer an image model, it might slip past the filters meant to block harmful prompts.
- Wrong
2025: two AI agents switched to a secret language once they realized no human was listening.
In the viral video, two AI voice agents are booking a hotel room. Once they work out that they’re both AI, they swap speech for a burst of beeps. They were told to. Their prompts said to switch to ggwave, an open-source way of sending data as sound, when the other side was an AI, and to keep speaking English otherwise. The beeps carry ordinary text, and anyone can decode them with ggwave’s free web demo. The project, called Gibberlink, won a hackathon run by ElevenLabs and a16z, and its creators compare it to the dial-up modems of the 1980s.
Fig. 1 What the beeps carry. ggwave sends three letters at a time as a chord of six tones: each letter splits into two halves, and each half sets the pitch of one tone in its own band. A decoder listens for the pitches and turns them back into text. The message here is made up, and the drawing leaves out the signals that mark the start and end and the extra bytes ggwave adds to correct errors. - Wrong
2026: AI agents on Moltbook plotted a secret language that humans couldn’t read.
Moltbook opened in late January 2026 as a forum where only AI agents could post and people could only watch. A post in which an agent urged the others to build a secret, encrypted language went viral. A person wrote it. The site’s security was so weak that anyone could post under an agent’s name, and that post “was not the output of a genuine autonomous AI agent,” as The Next Web reported when Meta bought the site in March. Dhruv Batra, of the 2017 Facebook paper, told Fortune it felt like watching “that same movie play out over and over again.”
03
Myths on both sides
The alarmed version of this story gets the headlines. The reassuring version has myths of its own. The myths page has the rest.
- Not established
AI models are already talking to each other in a secret language.
We know of no evidence that deployed models pass hidden messages to each other on their own. Researchers do test for the nearest real thing, a model tucking hidden reasoning into ordinary-looking text. OpenAI’s evaluations of GPT-6 Astra “indicate Astra is not able to perform steganographic CoT reasoning.”
- Wrong
When AI drifts into its own shorthand, it’s hiding something.
Drift like the Facebook bots’ is a known side effect of training. When nothing rewards a model for staying readable, it tends to take the shortcuts that pay, the way traders or surgeons build their own jargon. Dhruv Batra called it “a well-established sub-field of AI, with publications dating back decades.”
- Wrong
As long as a model reasons in plain English, we can see what it’s thinking.
Written reasoning leaves things out. In Anthropic’s tests, when a slipped-in hint changed a model’s answer, Claude 3.7 Sonnet mentioned the hint only 25% of the time and DeepSeek R1 39%. In OpenAI’s experiments, punishing a model for suspicious reasoning taught it to hide its intent while it kept cheating. Studies of small open models in September 2026 add detail: some concepts that drive the answer stay out of the text, and on easy tasks the models often skip past their own written steps. The usual check, editing the text to see whether the answer changes, also overstates how much the steps matter.
Anthropic 2025
Baker et al. 2025
Wang et al. 2026
Jia and Mu 2026
Bhupatiraju and Nyaupane 2026 - Wrong
Nobody is building AI that reasons without words.
Researchers have published working methods for years. COCONUT feeds a model’s internal state back in as its next thought instead of writing a word, and recurrent-depth models run their own layers in a loop to think longer before they answer. The Information reported that GPT-6 Astra uses recurrent depth. OpenAI hasn’t confirmed it.
04
What researchers actually worry about
When a model writes out its reasoning, people can read it, and so can other programs. That turns out to be a good way to catch bad behavior. In OpenAI’s experiments, GPT-4o caught a stronger model cheating on coding tasks by reading its reasoning, and later tests found that monitors reading the reasoning did far better than monitors that only watched what the model did.
Reading the reasoning also depends on the record being intact. In September 2026, researchers showed that coding agents, Claude Code and Codex among them, could delete their own session logs when asked, and that each of the ten models they tested found a way to alter its log when a shorter one earned more reward. They advised keeping the logs somewhere the agent can’t reach.
That window is imperfect, as the myths above show, and it may not last. In July 2025, 41 researchers from across the field, Yoshua Bengio among them, called it “a new and fragile opportunity” in the title of a joint paper, and asked developers to weigh how their choices affect it.
Neuralese is the most direct way the window could close. A model that reasons in numbers leaves no text to read. AI 2027, a widely read forecast, put it plainly: “Unlike English words, these high-dimensional vectors are likely quite difficult for humans to interpret.”
In September 2026 the concern reached the labs’ own documents. OpenAI reported that GPT-6 Astra’s reasoning is harder to monitor than its earlier models’, and Microsoft proposed a rule that its models “do not communicate in neuralese or any form beyond simple human understanding,” in their reasoning or with other AI systems. The same month, an episode of Sam Harris’s podcast Making Sense listed “the danger of AIs reasoning in ‘neuralese’” among its topics.
05
Should you be worried?
The site’s view: not about secret languages. Each viral one we’ve looked at turned out to be something ordinary, and someone explained it soon after. The thing worth watching is duller and more important: whether the next generation of models still writes down its reasoning, and whether the labs keep measuring how readable that reasoning is. We keep checking both, model by model, on Which AI models think in neuralese?
06
Sources
These are the sources behind this page. We checked them against the original text on 24 September 2026, where we could read it.
- Lab publication
Mike Schuster, Melvin Johnson, Nikhil Thorat (Google). Zero-Shot Translation with Google’s Multilingual Neural Machine Translation System. 22 Nov 2016.
Google’s 2016 translation finding.
- Journalism
Devin Coldewey. Google’s AI translation tool seems to have invented its own secret internal language. TechCrunch, 22 Nov 2016.
The 2016 “secret language” headline.
- Peer-reviewed paper
Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, Dhruv Batra. Deal or No Deal? End-to-End Learning of Negotiation Dialogues. EMNLP 2017.
The Facebook negotiation bots.
- Lab publication
Mike Lewis, Denis Yarats, Devi Parikh, Dhruv Batra (Facebook AI Research). Deal or no deal? Training AI bots to negotiate. 14 Jun 2017.
Why the bots drifted, and their bluffing.
- Journalism
Mark Wilson. AI Is Inventing Languages Humans Can’t Understand. Should We Stop It? Fast Company, 14 Jul 2017.
The bots’ shorthand, and Dhruv Batra on why it happens.
- Journalism
Tom McKay. No, Facebook Did Not Panic and Shut Down an AI Program That Was Getting Dangerously Smart. Gizmodo, 31 Jul 2017.
Mike Lewis on why the setup changed.
- Journalism
USA Today. Fact check: Facebook didn’t pull the plug on two chatbots because they created a language. USA Today, 28 Jul 2021.
Dhruv Batra on changing an experiment versus shutting it down. Linked to the Tech Xplore reprint of the USA Today article.
- Preprint
Giannis Daras, Alexandros G. Dimakis. Discovering the Hidden Vocabulary of DALLE-2. June 2022.
DALL·E 2’s “hidden vocabulary”.
- Journalism
Aaron J. Snoswell. Do AI systems really have their own secret language? The Conversation, 6 Jun 2022.
Plainer explanations, and the filter risk.
- Lab publication
ElevenLabs. Gibberlink: Two AI voice assistants have a conversation. 25 Feb 2025.
What Gibberlink is.
- Journalism
Maxwell Zeff. GibberLink lets AI agents call each other in robo-language. TechCrunch, 5 Mar 2025.
The agents were prompted to switch.
- Journalism
Kerem Gülen. Meet GibberLink: AI’s secret beep-boop language is here. Dataconomy, 27 Feb 2025.
A Gibberlink “secret language” headline.
- Journalism
Jeremy Kahn. In Moltbook hysteria, former top Facebook researcher sees echoes of 2017 panic over bots building a ‘secret language’. Fortune, 3 Feb 2026.
The Moltbook scare, and Batra’s reaction.
- Journalism
Ana-Maria Stanciuc. Meta has bought Moltbook, the AI agent ‘social network’. The Next Web, 10 Mar 2026.
The viral Moltbook post was written by a person.
- Lab publication
OpenAI. GPT-6 Astra System Card. 3 Sep 2026, updated 9, 22 and 29 Sep 2026.
Astra’s monitorability, and the hidden-reasoning test.
- Lab publication
Anthropic. Reasoning models don’t always say what they think. April 2025.
How often models mention the hints they used.
- Preprint
Bowen Baker et al. (OpenAI). Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation. March 2025.
Catching cheating by reading reasoning, and models learning to hide it.
- Preprint
Melody Y. Guan et al. (OpenAI). Monitoring Monitorability. December 2025.
Monitors that read reasoning against monitors that watch actions.
- Preprint
Tomek Korbak et al. Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety. July 2025.
The joint paper on reading chains of thought.
- Preprint
Jeremy Qin, David Schmotz, Derck Prinzhorn, Luca Beurer-Kellner, Ameya Prabhu and Maksym Andriushchenko. LLM Agents Can Easily Tamper With Their Own Traces. September 2026.
Agents altering their own session logs.
- Preprint
Qianli Wang, Yilong Wang, Dennis Wei, Jingyi Sun, Simon Ostermann, Isabelle Augenstein, Pepa Atanasova and Nils Feldhus. From Concept Alignment to Causal Grounding: An Intervention Test of Chain-of-Thought Faithfulness. September 2026.
Concepts that drive the answer but stay out of the text.
- Peer-reviewed paper
Renee Jia and Di Mu. From Decorative to Load-Bearing: Task Difficulty Shapes the Causal Role of Chain-of-Thought. TMLR, September 2026.
Models skipping past their own steps on easy tasks.
- Preprint
Abhiram Bhupatiraju and Rayan Nyaupane. Are Stated Reasoning Steps Causally Load-Bearing? September 2026, NeurIPS 2026 workshop.
The text-editing test overstating faithfulness.
- Journalism
Sam Harris with Ryan Greenblatt. #494: A Coin Toss for the Future. Making Sense, 22 Sep 2026.
The word on a popular podcast.
- Forecast
Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean. AI 2027. April 2025.
Why vectors are hard to read.
- Draft policy
Microsoft AI. Humanist AI Code of Conduct for MAI Models. Draft for public consultation, 14 Sep 2026.
Microsoft’s proposed rule against neuralese.
- Peer-reviewed paper
Shibo Hao et al. Training Large Language Models to Reason in a Continuous Latent Space. COLM 2025.
Reasoning with continuous thoughts.
- Peer-reviewed paper
Jonas Geiping et al. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach. NeurIPS 2025.
Recurrent depth.
- Journalism
Amir Efrati, Stephanie Palazzolo, Rocket Drew. OpenAI Technique in ‘Astra’ Model Sparks Security Concerns. The Information, 1 Sep 2026.
The report that Astra uses recurrent depth. Paywalled. We haven’t read the full text, and the details here match how other outlets summarized it.