Original title: GlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions
Elias Stengel-Eskin, Newton Sander, Carlos Bonetti, Sasha Boguraev +3 · September 2026
When artificial intelligence programs talk to each other, they might create their own ways of speaking. This creates a problem because humans need to monitor these systems to keep them safe. To study this, researchers built a platform called GlossoGen. They set up a scenario called SaveVeyru, where computer agents with partial information must talk to each other under pressure. The researchers found that these agents actually evolve new languages. The new languages build complex words and sentences, but they drift away from English until humans cannot understand them. This evolution requires pressure to be efficient, strong artificial intelligence models, and a stage where agents can agree on language rules. New agents can learn these languages just from seeing them used. Interestingly, while strong models are required to invent a language, weaker models can still learn it later. These artificial agents show signs of cultural evolution, which was previously seen only in humans.
Imagine a group of tourists in a foreign country who need to quickly share urgent messages about finding shelter, but they have no common language. They start by using gestures and simple English, but under pressure, they invent new words and rules that locals can't understand. In this paper, AI agents in a similar pressured scenario develop new languages that are efficient for them but opaque to humans, showing how communication evolves under constraints.
This matters because if AI agents develop their own languages, it could make them harder to monitor and control for safety. It also shows that AI systems might evolve culturally, which changes how we think about their capabilities and potential risks.
The abstract does not explicitly state limitations, but the study is based on a single scenario (SaveVeyru) and current LLMs, so results may not generalize to all multi-agent systems or future AI models. It also focuses on short-term evolution and does not address long-term safety or real-world linguistic applications in depth.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
GlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions — Elias Stengel-Eskin, Newton Sander, Carlos Bonetti et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.