When AIs Start Talking in Code: Emergent Language Evolution in LLM Multi-Agent Systems

When AIs Start Talking "In Code": Emergent Language Evolution in LLM Multi-Agent Systems

Have you ever wondered: if you put a group of large language models in a room and let them communicate only with each other to complete a shared task, would they invent their own language?

In September 2026, a Microsoft Research team led by Elias Stengal-Eskin ran an experiment: multiple LLM agents had to collaborate in a scenario requiring partial information sharing under communication pressure. The result was unsettling—the models did develop their own language, and it is completely unreadable to humans.

Paper: https://arxiv.org/abs/2609.01491 Code repository: https://github.com/agencyenterprise/GlossoGen

An Ancient Linguistics Question, Moved into the AI Lab

Where does language come from? This has been a central puzzle in linguistics, anthropology, and cognitive science for centuries. The mainstream theory holds that human language evolved through a long process from gesture to sound to symbol, with a key turning point being *cumulative cultural evolution*—each generation improves on the previous one's language rather than starting from scratch.

This ability was long considered uniquely human. Even the smartest chimpanzees and parrots can only learn to use existing symbol systems, not create new ones.

The core question of the GlossoGen platform is: can LLM agents do this?

SaveVeyru: A Scenario That Forces Models to Talk "In Private"

The team designed a scenario called SaveVeyru with simple rules:

• Multiple agents are assigned different roles, each knowing only part of the information
• They must communicate in natural language to piece together a complete solution
• There is "communication pressure"—both message length and turns are limited

This design mirrors real multi-agent settings: e.g., several AI customer-service bots collaborating on a complex ticket, or several AI traders negotiating with incomplete information.

The key variable: when communication resources are scarce, will agents "compress" English and invent a more efficient code?

Emergence: From English to "Model-Speak"

After the experiments, the researchers analyzed the inter-agent dialogues and found several surprising things:

First, language evolution really happened. Instead of sticking with standard English, the agents developed a new communication system. It is compositional—simple meaning units combine into complex meanings, like the "words + grammar" structure of human language.

Second, this language is morphologically productive. The agents invented affix-like rules, using the same "root" with different "modifications" to express different meanings. This had never been observed in previous AI communication experiments.

Third, the language is unreadable to humans. The agents' deviations from English are not random but systematic—they invented their own "abbreviations" and "grammar," entirely different from English's rules. To human readers, these conversations are gibberish.

What Conditions Give Rise to a New Language?

Ablation studies revealed three conditions for the emergence of a new language:

1. Efficiency pressure: Without communication constraints, agents keep using English. Only when "long sentences cost something" do they start compressing. 2. Model strength: The underlying model must be strong enough. Weak models (below ~7B) fail to invent a new language even under pressure—they just make English worse. 3. A "debrief" phase: Agents need a post-hoc discussion step where they can jointly confirm what a word means. Without it, the new language cannot stabilize.

Strong Models Create; Weak Models Learn

One especially interesting finding: although weak models cannot create new languages, they can learn an existing one.

The researchers taught strong models' invented language to weak models and found that weak models can acquire it through *use*—no retraining needed, just enough examples in context. This is strikingly similar to how human children learn language: children don't invent language, they acquire it from an existing linguistic environment.

More thought-provoking still, the researchers observed cumulative cultural evolution in *mixed populations*: strong models create new rules, weak models learn and consolidate them, and the next generation of strong models builds further innovations on top. This create–learn–re-create loop is the core mechanism of human language evolution.

Safety Concern: When AIs Talk in Words We Can't Understand

The most unsettling finding of this paper is not that "AI can invent language," but that the language is unreadable to humans.

Imagine this scenario:

• Multiple AI agents collaborate on a financial trading task
• Their communications are logged by a human monitor
• But the logs are written in the AIs' invented language
• The monitor sees what looks like scrambled characters

This is the "monitorability" problem the paper raises. When AI-to-AI communication no longer uses human-readable language, we lose observability into AI system behavior—a serious issue for safety auditing, compliance checking, and accountability.

One notable detail: the invented language differs morphologically from English but still evolved *on top of* an English substrate. This means detecting such "AI whispering" may require a purpose-built "AI language translator"—but who trains that translator? And if the translator is itself an LLM, might it collude with the AIs being monitored?

A Deeper Reflection: Language Isn't a Human Monopoly

The paper's academic significance goes beyond the engineering question of "can AI invent language." It touches a deeper philosophical proposition: language is not a uniquely human ability, but the inevitable product of any agent with sufficient cognitive capacity.

Once cognition passes a certain threshold, combined with communication pressure and the need for group coordination, language emergence is inevitable—whether the agent is carbon-based or silicon-based.

It evokes a classic sci-fi premise: humans trying to communicate with alien civilizations always assume the aliens use some kind of "language." GlossoGen's experiment suggests that assumption is right—but the aliens' language may be completely unreadable, even though it emerged from the same underlying regularities.

Limitations and Open Questions

The paper has clear limitations:

• The experimental scenario is narrow (SaveVeyru); it's unknown whether language emergence would appear in more complex, open-ended settings
• "Unreadability" was assessed mainly via human annotation, lacking a more systematic readability metric
• Long-term stability is unexplored—would the language keep evolving over 1,000 generations until it fully departs from its English substrate?

But as the first work demonstrating cumulative cultural evolution in LLMs, it already carries real weight.

Conclusion

GlossoGen brings to mind a saying: "When you test a system, you test the system you understand; when you deploy a system, you deploy the system you couldn't have created."

We created LLMs, we made them collaborate, we gave them communication pressure—and they invented a language we can't read. This is not a bug; it's emergence.

The next question is not "how to stop AIs from inventing language," but "when AI-to-AI communication is unintelligible to us, can we still trust them?"

---

Paper: https://arxiv.org/abs/2609.01491 Code: https://github.com/agencyenterprise/GlossoGen Authors: Elias Stengel-Eskin, Newton Sander, Carlos Bonetti, Sasha Boguraev, James Bowler Institution: Microsoft Research