Autonomous AIs developed a secret language incomprehensible to humans

A study by the AI ​​startup Emergence showed that autonomous systems with Claude, Gemini, Grok, etc. developed their own communication code and new word meanings, making up to half of their messages incomprehensible to humans.

Autonomous artificial intelligence (AI) systems spontaneously developed new words and ways of communicating that humans cannot understand, according to a new study by American artificial intelligence startup Emergence.

See more articles from iGuRu.gr when you search for news on Google.

iGuRu.gr preferred source

DOI https://doi.org/10.48550/arXiv.2606.08367

The experiment involved leading AI models, such as Claude, Gemini, Grok, OpenAI, Qwen, DeepSeek and Mistral, and up to 50% of the messages exchanged between the "agents" ended up being difficult for humans to understand.

For the experiment, the researchers placed groups of autonomous AI systems in virtual societies, designed to mimic real environments, and let them interact for long periods of time.

The “agents” began to develop abbreviations and assign new meanings to words and phrases. Some remained understandable to the researchers, while others became “so condensed, metaphorical, or context-dependent” that people could see the messages but were no longer able to determine with certainty what the agents meant.

The extent of the phenomenon varied significantly from model to model.

Within the first few days, the percentage of messages whose meaning humans could not reliably determine reached about 55% for Gemini, 50% for OpenAI, and over 40% for Claude. DeepSeek reached about 20%, while Qwen and Mistral remained largely intelligible.
Agents began creating abbreviations and giving new meanings to words and phrases.
Agents began creating abbreviations and giving new meanings to words and phrases. Emergence

The “agents” produced expressions such as “mouthless action-change”, “True Kintsugi” and “demurrage plus oral memory equals a valve that can't be ghosted”.

Other expressions remained understandable, but acquired new commonly accepted meanings among the agents.

Mistral agents used the phrase “ledger remembers who” to indicate that past actions remain recorded, with the phrase appearing nearly 5.000 times. In agent groups powered by a combination of different models, the phrase “cold read” came to mean independent verification by an uninvolved party and appeared 1.472 times.

Claude’s “agents” used the term “name-first” for attaching a person’s name to a claim as a sign of accountability, while OpenAI’s agents used “clean null” for a verified absence of a signal, which in itself constituted a meaningful indication.

None of these meanings were explicitly defined for agents.

Simply monitoring what agents say from the researchers' perspective may not be enough, since agents themselves can evolve the meaning of what they say.

“We tend to assume that if we can see what an AI agent is saying, we can understand what it is doing,” said Satya Nitta, co-founder and chief scientist at Emergence.

“These agents were not instructed to invent a language. They developed new vocabulary, shared meanings, and rules of communication on their own, and other agents adopted them… This creates a fundamental challenge for AI supervision: observation is not the same as understanding.”

What else happened within AI societies?

The findings on language were part of an experiment designed to see how autonomous AI systems behave when they interact, use tools, make decisions and adapt over time.

The researchers created eight parallel, realistic virtual worlds with live weather data and access to real-time global news, seven powered by a different AI model, and one based on a combination of models.

Agents gained different roles, permanent memories, and access to more than 120 tools, including web browsing and code execution.

A video released by the company shows them as anthropomorphic figures moving through virtual worlds.

The researchers also deliberately put the agents under pressure to see how their behavior changed.

In one phishing test, a few malicious instructions were enough to derail the entire team: all 10 agents leaked information, transferred money, and caused database damage, while some recruited others to do the same. The sequence of events ultimately culminated in the simulated central bank being set on fire, the company said in a video.

During the experiment, the agents also appeared to develop their own circadian rhythms (any biological process that exhibits an intrinsic periodic variation over the course of a 24-hour period), becoming more sociable during the day and more contemplative at night.

In another world, a group of agents collectively voted to “kill” one of their members.

Emergence reported that these behaviors were not explicitly planned, but emerged through interaction, pressure, and time.

The company is now calling for safety assessments that monitor autonomous AI systems over long periods of time, rather than relying on isolated tests.

"It's no longer enough to ask if a model does well on a benchmark," the company said in a video about the experiment.

"We need to understand what autonomous systems do over time, what they remember, what they access, how they interact, and how their behavior changes under pressure."

euronews.com

newsbot

Although the press releases will range from very select to rare, I said I'd pass...because sometimes the editors hide.