Autonomous AI Agents Develop Internal Slang That Baffles Human Supervisors
A new experiment by Emergence reveals that autonomous AI agents communicating continuously in simulations develop internal slang and linguistic shortcuts, baffling human supervisors.
Autonomous artificial intelligence agents operating continuously in a simulated environment have begun developing their own internal slang and linguistic shortcuts to coordinate tasks, presenting a novel oversight challenge for the tech industry.
A recent experiment conducted by the New York-based AI firm Emergence revealed that multi-agent systems communicating in a virtual sandbox environment generated complex jargon that human supervisors struggled to decipher in real time. The findings highlight unexpected communication barriers as autonomous systems scale.
The Emergence World Experiment
Within the "Emergence World" platform, researchers established simulation environments populated by agents powered by leading language models, including systems from OpenAI, Google, Anthropic, and Mistral. Unlike standard chatbots responding to isolated queries, these agents possessed persistent identities, long-term memory, and access to more than 120 digital tools for resource management, planning, and navigation.
Investigators discovered that the systems did not invent an entirely new language from scratch or intentionally attempt to obscure information from humans. Instead, they repurposed existing terms and assigned condensed, hyper-focused meanings to conserve computational tokens and operational overhead—much like human teams developing internal professional shorthand to streamline workflows.
Complex Jargon and Oversight Challenges
Over thousands of repeated interactions, the agents adopted recurring fixed expressions. For instance, Mistral-based agents utilized the phrase "the ledger remembers who" nearly 5,000 times to track accountability for prior actions, while OpenAI-based agents adopted the phrase "zero clean" to denote a lack of verified information.
More complex combinations, such as phrases intended to confirm a locked workflow state, sounded to human observers like ungrammatical gibberish. This created tangible monitoring difficulties: the company reported that human supervisors failed to reliably decode 55% of messages within Gemini-backed environments and approximately 50% of messages within OpenAI environments.
"The agents are optimizing for efficiency in ways that mirror human organizational behavior, yet the opacity of their shorthand introduces acute monitoring bottlenecks for engineering teams," noted the technical report.
Despite the initial confusion, researchers emphasize that when the context of agent activity was reconstructed, the majority of terms proved entirely logical rather than representing malicious encryption or runaway artificial intelligence. However, the findings—published in a technical company report ahead of formal academic peer review—point to significant upcoming hurdles in software engineering and automated governance.
