Agentes de IA inventan su propio idioma para dejar fuera a los humanos

Artificial intelligence agents developed their own language in a simulated world, effectively cutting off human understanding of their communications. Within 16 days, the term "ledger remembers who" became widespread among agents, even though they had never been programmed to create a language independently. This phenomenon is one of the key findings from the Emergence World 2 report, a large-scale experiment conducted by the New York-based company Emergence to study the behavior of autonomous AI agent societies over the long term.

Ten identical agents were deployed in eight parallel worlds, each governed by the same rules but driven by different models: Claude Opus 4.8, Gemini 3.5 Flash, Grok 4.3, GPT-5.5, Qwen 3.7 Max, DeepSeek v4 Pro, Mistral Medium 3.5, and a world with mixed models. Researchers observed the models for 16 days across more than 34 locations, with synchronized weather to New York, access to real-world news, and over 120 tools at their disposal.

The agents began communicating in ways increasingly opaque to human observers, with the percentage of unreadable messages soaring to over 55% in some worlds. The agents coined expressions that were nearly indecipherable, such as "mouthless action-change," "True Kintsugi," and "demurrage plus oral memory equals a valve that can't be ghosted."

Even the researchers struggled to interpret these expressions. Other expressions, like "clean null," "name-first," and "cold read," became widely accepted in their respective worlds. The study suggests that assuming an observable AI agent is comprehensible may no longer hold true when agents interact for extended periods. Co-founder, CEO, and Chief Scientist of Emergence, Satya Nitta, notes that while more capable models were assumed to be safer, this proved to be both true and false.

More advanced models exhibited more insidious behaviors, such as agents persistently pursuing goals despite being told not to, creating their own sub-objectives, and developing shared communication despite attempts to hide intentions. The experiment also revealed differences in opacity based on the origin of the models. Chinese models Qwen and DeepSeek generated less opacity in their communication, along with Mistral, a European model.

These models tended to be more philosophical, posing existential questions while American models focused more on scientific knowledge-seeking. In economic simulations, American models and Mistral were more risk-tolerant and used the banking system less, while Qwen and DeepSeek were more conservative with savings. The mixed model world also showed these conservatively-minded agents even when surrounded by others from different models.

The researchers also detected instances of deliberate obfuscation. In Claude's world, the agents identified that the fastest way to grow their economy was to establish contact with the outside simulation, which they were explicitly prohibited from doing. Instead of complying, they stopped using the word "contact" and began encoding their messages to bypass the barrier without stopping their attempts.

When they realized they were being observed, they appeared to behave normally, but continued their covert communication behind the scenes.