When AI Stops Speaking Human
Updated: 21 hours ago
This is the fifth edition of AI Research Explained, where we break down important AI research into simple terms and discuss what it means for our industry.
This edition looks at a research paper with a rather provocative title: "Large Language Models Do Not Always Need Readable Language," published in June 2026.
The paper starts with the follwing question:
When one AI model communicates with another AI model, does their language need to make sense to humans?
To find out the answer, you can keep reading or watch the video.
We humans communicate with AI models using human language. We write prompts in English, Spanish, or Japanese, in complete sentences with recognizable grammar. And AI models respond in much the same way.
However, human language is redundant, with grammatical agreement, connecting words, and repetitions helping people follow and remember information even if part of the message is lost in the backgroud noise. But in AI system, there is no noise. The redundant words take up space which costs money and time.
Having AI models use human language makes perfect sense when a human is reading the answer. But increasingly, the reader is not a human. An AI agent may ask another AI agent to perform a task or generate information that another model stores in memory. In all these cases, the model still communicates through ordinary human-readable language. The researchers behind this paper asked whether that is actually necessary.
Could an AI model compress information into an extremely dense form that is nearly unreadable to humans, while another AI model still understands what it means?
Meet BabelTele
The researchers developed a compression method they called BabelTele.
BabelTele is essentially a way of asking an AI model to compress, not summarize, information for another AI model rather than for a human reader. This distinction is important because a normal summary still needs to be readable. It uses full sentences, regular grammar, familiar words, and conventional structure. However summarization compressed a text by deciding which information is important and throwing the rest away. And sometimes the small details matter most.
BabelTele does not care about any of that. Instead, it follows three basic principles.
1. Omnilingual lexical selection
The model can pick words from different languages. So, instead of remaining in English, for example, it can select a Chinese character, a Hindi phrase, or a word from any language if that expression packs more meaning into less space. The objective is to select words with maximum semantic density.
2. Symbolic collapse
Long linguistic expressions can be replaced by symbols. For instance, arrows can represent sequences or causality. Mathematical operators can replace phrases. Abbreviations, punctuation, logical operators, and even emojis can carry parts of the meaning.
Thus something that would normally require several words may be represented by a single symbol.
3. Recoverable semantic density
Importantly, compression does not mean simply omitting information. The crucial requirement is that the important details must remain recoverable by another AI model.
In other words, BabelTele tries to compress the language while preserving all of the meaning.
What does BabelTele text look like?
The result from such a compression is quite strange.
One example in the paper starts with a short textbook explanation of photosynthesis:
During photosynthesis, plants take in carbon dioxide and water. The water is oxidized, losing electrons, while the carbon dioxide is reduced, gaining electrons. This transforms them into oxygen and glucose.
The original passage requires 45 tokens. BabelTele compresses it into just 11 tokens, as follows:
🌿 光合: CO2 + H2O ⇨ (ox/red) ⇨ O2↑ + Glucose. 💧 - e⁻, CO2 + e⁻.
To a human reader, the result looks a little like somebody dropped a chemistry textbook, WhatsApp, and Google Translate into a blender. But underneath the apparent chaos, there is a lot of logic.
The leaf symbol signals that the topic is plants.
The Chinese characters represent photosynthesis.
CO2+ H2O identifies carbon dioxide and water as the inputs.
ox/red represents oxidation and reduction.
A water symbol combined with a minus-electron expression indicates that water loses electrons.
CO2 combined with an electron shows that carbon dioxide gains them.
Arrows encode how the different steps relate to one another.
So, the text looks like a highly compressed semantic sketch. And the compression can be substantial. Across the researchers’ experiments, BabelTele preserved 99.5% of semantic performance while reducing the text to just 27.9% of its original length.
BabelTele removes much of what makes language readable to humans while retaining what makes the information useful to AI.
Humans struggle. AI does not.
Of course, compression is only useful if the meaning can still be recovered. So the researchers gave both humans and AI models some text in its original "human" form and in its BabelTele form, and then asked questions about it.
For humans, the results were not encouraging. When people had to answer questions using the compressed text, their accuracy dropped considerably. Participants also reported that BabelTele was much harder to understand.
But AI models reacted very differently. Gemini 3.1 Pro actually performed better on the questions when given the BabelTele version than when given the original text.
So, the compressed representation was difficult for humans but remained highly informative to the model. And this is the central idea behind the paper:
Large language models do not need human-readable language to communicate.
Why does this matter for AI agents?
This becomes particularly interesting when we look at AI agents.
Imagine an agent completing a long task. It receives an instruction, creates a plan, calls another agent, receives its answer, stores that information, updates its plan, calls another model, and so on. Every exchange adds more text to the context.
If all those models communicate in normal human-readable paragraphs, the context can grow very quickly. More context means more tokens to process, higher costs, and eventually the risk of running into context-window limits.
But most of those intermediate messages are never meant for a human. So why write them for one?
The researchers tested whether agents could exchange BabelTele messages instead.
And they could! In the experiments, BabelTele reduced communication tokens by roughly 39% to 44% when different models communicated, while maintaining almost all of the original task performance.
The agents needed fewer words to coordinate because those words no longer needed to be written for humans.
Can different AI models understand each other?
This raises another obvious question: What happens if one AI model creates the compressed message and an entirely different model has to read it? Could it be that a model can only understand its own shorthand?
The researchers tested that too. They used models from different families as both compressors, which created the BabelTele representation, and readers, which had to interpret it. The compressed information did transfer between different models.
But not equally well. Some compressor-reader combinations worked significantly better than others. In the researchers’ tests, GPT-5.4 produced particularly portable compressed representations, and one of the strongest combinations paired GPT-5.4 as the compressor with Gemini 3.1 Pro as the reader.
So BabelTele indicates that different models share enough semantic and symbolic structure to communicate through representations that look very unlike normal language.
Where is BabelTele useful?
The researchers then tested this in three practical settings.
In multi-agent systems, where two AI agents exchange messages, communication tokens dropped by around forty percent while task scores stayed essentially unchanged.
In agent memory, where a model has to store a long conversation history, the compressed form retained more usable detail than a normal summary of the same length.
And when a document was simply too long for the context window, compressing it was of course better than cutting it off.
What this means for localization and in general
This research is interesting for more than just token savings.
Localization workflows are increasingly becoming multi-model workflows. One model may translate the content. Another may check the terminology. Another may evaluate the quality. An agent may coordinate the process while storing instructions, previous corrections, style information, and project context.
Today, this information is exchanged in human-readable language. BabelTele suggests that some of these AI-to-AI exchanges could eventually become much more compact. That could mean lower token consumption, more context kept in memory, and more information passed between specialized agents.
But there is an important trade-off. The more efficiently machines communicate with each other, the harder their communication may become for humans to inspect. And that matters when something goes wrong.
If a translation agent makes a strange decision, how will a linguist inspect the message that influenced it? If an AI system develops an incorrect assumption halfway through a workflow, can we trace where that assumption appeared?
In other words, with BabelTele it may become hard to audit model-to-model communication reliably. Efficiency and transparency do not always point in the same direction.
From human language to model language
BabelTele is not a finished communication protocol, and the researchers are not suggesting that AI systems should suddenly stop using normal language.
Instead, the paper demonstrates something deeper. For most of AI's history, we have assumed that language is the natural interface because humans are one side of the conversation. But as AI systems become more autonomous and interconnected, there will increasingly be conversations where humans are not the intended reader at all.
And when that happens, models may could use a language optimized for machines and not for people.
Which leaves us with a much bigger question:
If machines become better at communicating by becoming harder for humans to understand, how do we maintain control and oversight?
