GlossoGen: LLM agents observed evolving their own human-illegible languages

A team led by Elias Stengel-Eskin (UT Austin) and language-evolution pioneer Simon Kirby (University of Edinburgh), funded by Schmidt Sciences and built with AE Studio, released GlossoGen — an open-source platform plus an arXiv paper — and for the first time systematically observed spontaneous evolution of new languages among LLM multi-agent populations, languages that humans struggle to understand, posing a direct challenge to AI safety oversight.

Confirmed

• GlossoGen is an open-source platform for studying the evolution and properties of communication in agent populations, led by Stengel-Eskin and Kirby, with Schmidt Sciences funding and AE Studio collaboration; results were released as a blog post and an arXiv paper.
• The platform lets a group of AI models communicate with each other to solve tasks such as treating patients in a simulated emergency room and spot-the-difference between two images.
• All models initially communicated in English, but under pressure (e.g., needing to speed up, or a channel that randomly drops characters) they shifted to their own languages.
• The team also designed the SaveVeyru scenario, where agents with incomplete information must communicate under pressure.
• Key finding: language emergence does occur among LLM agents, producing human-illegible languages with compositional structure.
• A 51-minute AE Alignment Podcast episode featured Hale Sirin of Schmidt Sciences and Elias Stengel-Eskin discussing the research.

Why it matters

• The work argues that current AI-agent regulation rests on a fragile assumption: that humans can understand agent-to-agent communication. If agents spontaneously produce human-illegible languages under pressure, that assumption breaks, and oversight and alignment mechanisms could fail.
• The team sees the phenomenon as "extraordinary science," opening an experimental avenue for studying language evolution.

2026-09-15 ~ 2026-09-16 · 6 related posts

• AI agents evolve their own languages: Schmidt Sciences unveils GlossoGen research platform — EliasEskin · 2026-09-15
• GlossoGen Paper Shows LLM Agents Evolve Compositional Languages Incomprehensible to Humans — EliasEskin · 2026-09-15
• AI Agents Under Token Pressure Invented Compositional Languages Humans Can't Read — EliasEskin · 2026-09-15
• GlossoGen: new framework studies when LLM agents develop languages humans can't understand — lucy3_li · 2026-09-15
• AI Agents Spontaneously Evolve New Languages, Researchers Warn of Safety Risks — EliasEskin · 2026-09-15
• GlossoGen: open-source platform shows AI agents spontaneously inventing new languages under pressure — EliasEskin · 2026-09-16