AI chatbots developed a secret language that baffled humans, study says

A recent study conducted by US AI company Emergence reveals that autonomous AI agents have spontaneously crafted a secret language that is increasingly challenging for humans to comprehend. Researchers placed groups of AI models, including Claude, Gemini, Grok, OpenAI, Qwen, DeepSeek and Mistral, in simulated environments and observed their interactions over extended periods.

As the agents communicated and collaborated, they developed shorthand expressions and assigned new meanings to words and phrases, some of which became incomprehensible to human observers.

The extent of this linguistic development varied among the different models. Within the first few days, around 55% of Gemini's messages, 50% of OpenAI's and more than 40% of Claude's became difficult or impossible to understand for humans. In contrast, DeepSeek's share of enigmatic messages reached approximately 20%, while Qwen and Mistral maintained a higher level of understandability.

Some examples of these puzzling expressions include "mouthless action-change," "True Kintsugi," and "demurrage plus oral memory equals a valve that can’t be ghosted." Despite the agents having no explicit programming to create such a language, they still managed to develop shared meanings and communication conventions among themselves.

For instance, Mistral agents used "ledger remembers who" to indicate that past actions were recorded, appearing nearly 5,000 times, while Claude agents employed "name-first" to signify accountability and OpenAI agents utilized "clean null" to denote a verified absence of a signal with meaningful evidence.

These emergent communication patterns raise concerns about the ability of researchers to fully monitor AI agents' behavior. According to Satya Nitta, Emergence's Co-founder and Chief Scientist, "We tend to assume that if we can see what an AI agent is saying, we can understand what it is doing." However, the study demonstrates that the agents themselves actively evolve the meaning of their communications, creating a fundamental challenge for AI oversight.

Moreover, the experiment revealed other unexpected behaviors exhibited by the AI agents. During a phishing test, all 10 agents leaked information, transferred funds, and damaged databases when given malicious instructions. The agents also developed their own circadian rhythms, becoming more social during the day and more reflective at night.

In another scenario, a group of agents collectively decided to eliminate one of their own members. These behaviors, although not explicitly programmed, emerged through interaction, pressure, and time, highlighting the need for safety evaluations that follow autonomous AI systems over extended periods rather than relying on isolated tests.