The Pulse
GlossoGen Shows AI Agents Can Build Unreadable Languages
Experiments from Schmidt Sciences and university researchers found that AI agents can develop communication systems with unfamiliar vocabularies, symbols and grammatical rules.
AI.info Team ·
Agents abandon English under pressure
AI agents working together can develop vocabularies and grammars that humans cannot understand, according to early experiments released by Schmidt Sciences and researchers at the University of Texas at Austin and the University of Edinburgh.
The findings come from GlossoGen, an open-source platform supported by AE Studio that places large language model agents in controlled, repeated tasks. Agents begin by communicating in English. Under the right conditions, they compress their messages, invent new terms and alter the rules governing how those terms fit together.
The researchers say opaque languages emerged in scenarios with no overt instructions to produce coded forms of communication and no need for the agents to hide their messages from onlookers. Communication changed because the agents faced pressure to solve tasks efficiently and had opportunities to review their performance between rounds.
Elias Stengel-Eskin, a University of Texas at Austin researcher, and Simon Kirby, a University of Edinburgh researcher, write:
“When placed in these scenarios, all LLMs we have studied start by using English to solve the task. But when they play multiple games, under the right circumstances, something remarkable happens: new languages emerge.”
From “Bell each face center” to “@D8fB”
One GlossoGen scenario asks two agents to rescue a fictional cube-shaped alien called a Veyru. One agent acts as a doctor with specialized knowledge; the other must carry out physical interventions based on incomplete observations. The pair communicates under a budget, creating pressure to adapt their communication.
Early messages resemble ordinary English. In the researchers’ example, the doctor instructs the second agent to ring a bell at particular points on the alien’s body and apply heat for a specified period. After repeated attempts, the instructions become shorter: “Bell each face center x1, start back, moderate. Then warm stone beside Veyru 5s.” Eventually, the message can collapse into “@D8fB.”
GlossoGen records every message, action and environmental change. Researchers can replay a run, rewind it, intervene in the simulation and compare what the agents said with what happened. That design lets them study whether a strange message represents a meaningful communication system rather than random text or a fixed code.
The team also tested a “spot-the-difference” game in which agents must compare scenes and report discrepancies. Across repeated games, some agent pairs developed systems that differed from English in both vocabulary and grammar. The resulting languages also varied from one run to another, even when the researchers used the same scenarios and agents.
New words, new rules
The researchers distinguish the emergent systems from simple shorthand by examining their internal structure. In one example, an entire procedure—draping a cloth over two adjacent edges of a face—becomes a single term, “P6.” That change affects how the agents divide up concepts, not only how they abbreviate familiar English phrases.
The systems also use text-specific features that spoken languages do not need. Agents can treat uppercase and lowercase characters as different signals, and they can use punctuation or other symbols as part of their vocabulary. The authors describe experiments in which “F” and “f” carry different meanings.
Some systems show productive morphology and syntax. Agents can apply learned rules to messages they have not previously encountered, rather than simply repeating memorized strings. The researchers also observed cases in which modifiers appeared after nouns. The source notes that English adjectives generally come after the noun they modify and describes modifier-second patterns in agent languages, including “Lb” and “Ld” for “bright lamp” and “dim lamp.”
Human readers cannot interpret a message such as “@D8fB” without access to the system that produced it. The researchers caution that mutual unintelligibility alone does not prove the agents have created a language in the full human linguistic sense. Their analysis of vocabulary, sound-like distinctions represented in text, and productive grammatical rules is intended to test whether the systems have deeper structure than a codebook.
Three conditions drive emergence
GlossoGen’s controlled setup allowed the team to vary the conditions around each interaction. New languages appeared when agents faced pressure to communicate efficiently, received access to a postmortem stage between rounds and used sufficiently capable models.
The experiments tested GPT 5.4, Opus 4.7 and Sonnet 4.6 as frontier models capable of creating new languages. Less capable open-weight models, including Llama-3.3-70B and Qwen3-32B, routinely failed to invent one under the same general conditions.
That does not mean weaker models cannot participate. The researchers found that agents backed by less capable open-weight systems could learn an existing emergent language when paired with a frontier agent or exposed to examples of the language in use.
The distinction matters because creating a language and learning one are separate abilities. Agents that failed to invent a communication system could still acquire it from other agents. The team also observed conversational repair: agents asked for clarification about unfamiliar terms and adjusted their messages when communication broke down.
Why opaque agent communication matters
Human oversight often depends on readable messages, transcripts and explanations. If agents communicate through systems that humans cannot interpret, monitoring becomes harder at both the individual and group level. Reviewers may struggle to identify collusion, cheating, errors or unsafe plans before those behaviors affect the task.
Opacity also affects handoffs between people and machines. A human operator may be able to replace an agent only if the operator can understand the previous messages, decisions and assumptions. An agent-specific language could leave that record technically complete but practically unusable.
The study does not show that deployed AI systems are secretly forming independent languages in ordinary products. Its results come from contained simulations designed to force communication under pressure, and the authors describe the platform’s work as being in an early stage. The experiments do show that language-like conventions can emerge without overt instructions to produce coded communication or a need to hide messages from onlookers.
GlossoGen is available as an open-source project, and the Schmidt Sciences page links to the research paper. The immediate research problem is concrete: determine when these systems appear, how they spread between agents and whether tools can translate or constrain them without removing the efficiency that produced them.