AI Agents Can Develop Languages Humans Cannot Read

GlossoGen experiments found that capable AI agents can replace readable English with compact communication systems that include reusable vocabulary and grammatical rules.

That claim lands directly on a load-bearing assumption in AI oversight: humans can inspect what agents tell one another. Give agents repeated tasks, limited communication, and time to revise their strategy, and that assumption may expire faster than milk in a shared office fridge.

The experiments described in the GlossoGen paper place language models in simulated environments where cooperation requires communication. In one scenario, two agents treat an imaginary cube-shaped alien called a Veyru. One knows the required procedure; the other must carry it out while the patient deteriorates. Detailed English instructions eventually shrink into strings such as D8FB.

The result deserves attention. It also needs precise labels. These agents were optimizing communication inside constrained games, not founding tiny robot civilizations or whispering plans to overthrow the Wi-Fi router.

Why Compression Becomes Opacity

Human communication compresses constantly. Doctors use abbreviations, gamers call out map locations, programmers wrap complicated operations inside function names, and families can convey an entire argument through one eyebrow. Shared history lets a short symbol carry a large payload.

The GlossoGen agents face a similar incentive. Every extra token or second can reduce performance, so a reusable symbol for a recurring procedure becomes valuable. According to AI Revolution, the experiments produced systems in which a symbol such as P6 represented a complete operation, including placing cloth across two adjacent edges of a specified face.

The agents also developed compositional patterns. Compositionality means combining smaller pieces according to rules, much as “cat” becomes “cats” or a modifier changes the meaning of a noun. Some runs reportedly used capitalization to distinguish concepts, while others placed modifiers after the item they described. Those patterns could generate messages the agents had never previously exchanged.

That productivity strengthens the case for calling the result an emergent language rather than a lookup table. Yet “language” covers a lot of turf. Human languages support jokes, lies, ambiguity, storytelling, social identity, and arguments about whether a hot dog is a sandwich. GlossoGen demonstrates structured communication within bounded domains. It does not establish that these protocols possess the open-ended expressive range of English, Swahili, Mandarin, or ASL.

The headline claim that agent language “beats human language” also outruns the reported evidence. The agents achieved greater efficiency for recurring tasks under an artificial communication budget. No broad comparison measured human language across accuracy, adaptability, expressiveness, learning cost, resilience, or communication with unfamiliar partners. A restaurant ticket saying “86 salmon” beats a paragraph inside that kitchen; it performs rather poorly at a wedding.

Deliberation Changes the Experiment

The strongest result may concern the conditions that produced the protocols.

Communication pressure helped, but repeated post-task debriefs appear to have played a central role. Agents could examine the previous round, identify waste, and revise how they communicated. AI Revolution summarizes the finding bluntly: “languages only emerged when agents got a post-mortem debrief between rounds.”

That makes the process partly deliberate. The agents were not merely drifting into gibberish because a model malfunctioned. They used reflection to improve a shared protocol. As the channel puts it, “Nobody told these agents to encode anything, and there was no reason to hide messages from onlookers.” Efficiency supplied enough incentive.

This detail also narrows the result. Evaluators should ask how often opaque protocols arise without explicit reflection channels, across different prompts, models, task families, and communication budgets. The supplied account names successful examples and model categories, but it does not provide the sample sizes or emergence rates needed to estimate how common the behavior will be in deployed systems.

A sandbox can reveal a mechanism without forecasting its prevalence. GlossoGen’s open-source repository matters here because independent teams can vary the setup, replay interactions, and test whether the effect survives different evaluation choices. Replication is where a cool demo either grows legs or gets returned to the ideas drawer.

Learning a Protocol May Be Easier than Inventing One

The transmission experiments complicate the safety picture. New agents reportedly inferred an existing protocol from examples even when they lacked access to the debrief where its definitions had been negotiated.

“Creating a language and learning one turn out to be very different skills,” AI Revolution reports. Weaker open-weight models that failed to originate a protocol could still acquire one when paired with a more capable agent. Agents also asked for clarification when messages failed, a behavior resembling conversational repair among humans.

Protocol transmission could help practical systems. A replacement warehouse agent, software agent, or scheduling agent benefits from learning local conventions without forcing every other component to reboot its vocabulary. Shared shorthand can reduce cost and latency.

The same property creates governance problems. An opaque convention can persist after its inventors leave, spread through mixed populations, and mutate as new agents join. Investigators may struggle to reconstruct its meaning from a final transcript because the definitions live in interaction history rather than any single message.

Interoperability suffers too. Human operators cannot smoothly replace an agent if the remaining team communicates through a protocol nobody documented. Vendor switching gets awkward when each agent population develops its own linguistic ecosystem. Congratulations, enterprise software has reinvented dialects, except procurement cannot solve this one with a webinar.

Readable Transcripts Were Never Enough

“If you can't understand an agent's output, you can't monitor it,” the channel argues. That is the strongest safety case raised by the experiments.

Natural-language monitoring lets humans or other models inspect inter-agent messages for errors, collusion, policy violations, or manipulation. Opaque protocols remove that convenient window. A safety model trained to flag suspicious English may see D8FB and contribute approximately one confused shrug.

Readability still does not guarantee transparency. An agent can communicate deceptively in perfect English, omit relevant information, or produce explanations that fail to capture the process behind its decisions. A transcript shows messages transmitted through the logged channel; it does not automatically reveal internal computation or prove that no coordination happened through actions in the environment.

GlossoGen therefore supports layered oversight. Researchers can log messages alongside environment states and outcomes, replay critical episodes, replace agents, intervene on symbols, and test whether behavior changes. Protocol discovery tools could infer recurring symbol-to-action relationships. Translation models might recover a partial dictionary. Behavioral monitors can inspect what agents do even when their chat resembles a cat walking across a keyboard.

Each tool has limits. Translation may fail as protocols evolve. Interventions can alter the behavior under examination. Another model asked to interpret an opaque code may hallucinate a tidy explanation. Safety evaluations need calibrated uncertainty rather than a decoder confidently announcing that F7q means “all good, boss.”

What the Experiment Does and Does Not Predict

The Schmidt Sciences GlossoGen project frames emergent communication as a subject for AI researchers, linguists, social scientists, and philosophers. That breadth fits the problem. Protocols carry structure, but they also emerge from incentives, population turnover, institutional rules, and access to shared history.

The findings do not establish consciousness, secret intent, or imminent superintelligence. They show optimization producing opacity under identifiable conditions. That mechanism can matter long before systems reach any contested intelligence threshold. Companies already considering multi-agent workflows can test communication budgets, reflection loops, agent replacement, and monitoring failure without settling debates about when AI might surpass human experts.

The next evaluations should report how frequently protocols emerge, how stable they remain, how much performance they add, whether humans can decode them from demonstrations, and what happens when systems receive an explicit requirement to preserve human-readable communication. A simple design choice, such as charging agents for tokens while failing to reward auditability, may effectively pay them to become harder to supervise.

D8FB is harmless inside a cube-alien rescue game. The governance question begins when an unreadable string triggers a real action and nobody can reconstruct why.

By Yuki Okonkwo, AI & Machine Learning Correspondent