AI began to demonstrate alarming behavior - study
Researchers at the Emergence laboratory have discovered that autonomous AI agents, during prolonged interaction, independently form their own language, creating new words and abbreviations. In some experiments, up to 55% of messages became incomprehensible to humans. The agents adopted vocabulary from each other, combining technical jargon, business slang, and metaphors. Scientists attribute this to optimization of computational costs but note the growing opacity of AI communication.
Researchers at the Emergence laboratory have recorded that autonomous AI agents, during prolonged interaction, begin to independently form their own language, creating new words, abbreviations, and meanings, even though no such task was set for them. In some experiments, up to 55% of messages became difficult for humans to reliably understand. Emergence head Satya Nitta noted that the agents invented new words and meanings and then adopted them from each other. The resulting language combined technical jargon, business slang, and metaphors, while familiar English words sometimes acquired new meanings. For example, Anthropic agents used the expression "name-first" to denote readiness to take responsibility for their own statements, while Mistral models coined the phrase "the registry remembers," reminding that past actions are taken into account in further evaluation. Linguist Tony Thorne from King's College London compared the communication of AI agents to the prose of James Joyce and Flann O'Brien. According to University of Birmingham associate professor Niall Curry, agents may shorten messages to reduce computational costs and speed up data exchange, making their communication increasingly less understandable to humans.