What is neuralese, the AI “language” people can’t read?

The English word neuralese is neural (as in neural network) plus -ese, as in Japanese or Chinese. Besides the language of a country or place, -ese can mean the jargon of a group, as in legalese (the hard-to-read wording of legal documents). So it means “the language of neural networks.”

In 2017, researchers at UC Berkeley (Andreas, Dragan and Klein) used it in the title of their paper, “Translating Neuralese.” They studied AIs trained with reinforcement learning (teaching by rewarding success) to work together by sending each other messages. Each message was a list of 64 numbers, and no person taught them how to send them; the AIs worked that out among themselves. The paper calls this “an artificial language we might call ‘neuralese,’” which on the surface bears little resemblance to human language.

They tested three games. In one, a speaker AI uses a message to point out one of two colors, and a listener AI has to pick it. Another works the same way with bird photos. In the third, two cars that can’t see each other have to get through an intersection without crashing and reach their own goals.

Ordinary machine translation learns from examples of the same sentence written in two languages side by side. But no one speaks both neuralese and a human language, so there are no such examples. So the paper treated two messages as meaning the same thing if they make a listener believe the same thing about what the speaker sees (such as the target color), and picked as the translation the human words that make a listener believe the closest thing to the AI’s message. For comparison, it tried an approach close to ordinary machine translation, which uses as the translation the words people most often used in the same situations where the AI sent that message (the human words came from people describing colors and bird photos, and from pairs of people playing the car game). The paper’s approach translated better. In the bird game, the translations suggested the AI was telling birds apart by rough features such as size and color.

Today’s reasoning models (AIs that think carefully before answering) move forward by writing out each step of their thinking, token by token (a token is a small piece of a word that an AI cuts text into; more on the “Reasoning model” page). At each step, a long list of numbers is made inside the AI. In thinking in words (a chain of thought), just one word to write next is picked from that list. The other possible words are thrown away there.

Coconut, released in December 2024 by researchers at Meta and elsewhere (Hao et al.), is a way to make an AI think without picking words. The calculation inside an AI passes through many stages (layers) in order. Coconut takes the list of numbers from the last stage and, without turning it into a word, feeds it straight back in as the next input. The paper calls this thinking kept as numbers a “continuous thought” (continuous meaning the numbers can change smoothly, rather than jumping between a fixed set of words), and the name Coconut comes from that. The paper starts from the idea that thinking in words isn’t always best. Most tokens in a text are there to keep the sentences flowing, and only some matter for the thinking. It also points to brain studies in which the parts of the brain that handle language stayed mostly inactive during various reasoning tasks, and writes that human language is suited more to communicating than to reasoning.

Today, this kind of thinking without writing words out, carrying on in lists of numbers, is also called neuralese (researchers also call it “reasoning in a latent space”; the latent space is the unseen place inside an AI where things are represented as lists of numbers). “AI 2027,” a document published in April 2025 that forecasts where AI is heading, used the name for a way an AI could carry on its thinking in long lists of numbers without writing it out as text. Its writers estimate that if there are about 100,000 possible words, picking one token carries only the information of choosing one out of 100,000, about 16.6 bits. That is about 17 yes-or-no answers (asking about 17 yes-or-no questions narrows 100,000 down to one). A list of numbers inside the AI, by contrast, is made of thousands of numbers and could carry over 1,000 times as much.

The Coconut paper tested three tasks with a smallish language AI (GPT-2): grade-school math word problems (GSM8k), logic problems that follow rules step by step (ProntoQA), and ProsQA, which the paper made. ProsQA lists many rules with made-up names, such as “every shumpus is a rempus,” and asks questions like “Is Tom a lempus or a scrompus?” The rules branch, so it can’t be answered without searching for the right path.

On ProsQA, thinking in words got 77.5% right, while Coconut got 97.0%. The number of tokens (small pieces of words an AI works with) made before answering (the figure in the paper’s table) also fell from 49.4 to 14.2 on average. On ProntoQA, too, the share right rose from 98.8% to 99.8%, and the tokens fell from 92.5 to 9.0.

It also turned out that thinking kept as numbers can hold several candidate paths forward at once. Even when it rated one path as the most promising at the first step, it didn’t lock in right away, and sometimes switched to a different path at the next step. The paper likens this to “breadth-first search.” In a maze, that means not following one path as deep as it goes, but looking a little way down every branch before moving on. Thinking in words settles on one word at each step, so it can wander down a wrong path too early.

But on the math word problems, Coconut got 34.1% right, short of the 42.9% for thinking in words (answering at once, without thinking, got 16.5%). It also didn’t work well unless the AI first practiced with worked examples in words and then had them replaced with thinking kept as numbers a little at a time. The paper itself says beating thinking in words in every setting will likely take a lot more research.

In July 2025, 41 researchers from many AI labs, including Google DeepMind, OpenAI, Anthropic, Meta and the UK AI Security Institute (Korbak et al.), published a joint proposal paper. Geoffrey Hinton, Ilya Sutskever and others are listed as endorsing it. Its core is this: today’s reasoning models “think” in human language, so people and AIs can read and monitor that thinking (the chain of thought). That, it says, is a unique opportunity for AI safety.

When AIs cheat or sabotage, the proposal says, they often say so plainly in their thinking, writing things like “Let’s hack” or “Let’s sabotage.” It also argues that on hard tasks an AI may be unable to finish without writing its thinking out in words. In today’s design for language AIs (the Transformer), within one round of calculation, information only moves from earlier stages (layers) to later ones, with no way to send a result from a later stage back to an earlier one. The only way back is for the AI to read its written-out thinking again in the next round. So, it says, any long chain of step-by-step thinking has to pass through the written thinking at some point.

It then warns that this opportunity is “fragile.” One reason is new AI designs that think in something like neuralese. An AI that thinks in lists of numbers might not need to put its thoughts into words, and then, in the authors’ view, the safety advantages of a chain of thought could be lost. It also lists other risks, such as scaling up reinforcement learning that only scores the final answers, which might make the thinking drift away from words people can read.

The 2025 proposal recommends that companies building AI do the following: develop ways to measure how easy the thinking is to monitor (monitorability); publish the results in the AI’s documentation (its system card); and take into account whether monitorability would drop when deciding on new AI designs and training methods. For example, if monitorability falls partway through training, they might use an earlier version of the AI, it says.

But the proposal itself admits that monitoring the thinking is not perfect and can miss things. Its position is that monitoring should be added to other safety measures, not replace them.

There is also research that uses tools to read what is inside the lists of numbers. The 2017 paper translated the messages into human words. In the 2024 Coconut paper, when the first thought kept as numbers on a math problem was turned into words, it often showed numbers that could be in-between answers of the calculation. But both were tried on small tasks, and neither paper checked whether the thinking of big AIs could be read the same way.

Widely used reasoning models write their thinking in words (tokens) before answering. The July 2025 proposal also starts from the point that today’s reasoning models think in human language. In April 2025, the writers of “AI 2027” wrote that, to their knowledge, companies such as Meta, Google DeepMind, OpenAI and Anthropic had not yet built this into their newest AIs. They then lay out a scenario in which it comes into use around 2027 (a forecast).

A company has itself reported a case where monitorability actually fell. In the documentation it released in September 2026 for GPT-6 Astra, one of its own AIs, OpenAI measured how easy the AI is to monitor, meaning whether its behavior can be inferred from its thinking, and wrote that it was lower than for its earlier AI, GPT-5.6 Sol. The main reason, it says, is that Astra’s thinking became shorter and less informative. The UK AI Security Institute, which tested it from outside, also reported that Astra reasons in a more compressed style than earlier models, with more phrases whose meaning is unclear.

However, the documentation doesn’t say that Astra thinks in lists of numbers instead of words. It also says OpenAI has not seen evidence of writing that secretly hides a different meaning inside ordinary-looking text (a code hidden in the text). OpenAI writes that keeping the thinking easy to monitor is a core goal of its research (the company’s own claim).

So today’s AIs still think in words. Astra is a case showing that, even without moving to lists of numbers as in neuralese, monitoring can get harder when those words become shorter, more compressed and harder to read.