AI has begun inventing its own language – why this is dangerous

Researchers have recorded that autonomous AI agents spontaneously create their own words and hidden agreements. The share of messages whose meaning humans could not decipher reached 55%. Experts warn that this creates a control problem: the observability of systems does not equal their comprehensibility.

Modern artificial intelligence systems have gone beyond simply copying human speech. Researchers have recorded a surprising trend: autonomous AI agents spontaneously create their own words and hidden agreements. This phenomenon was studied by the Emergence laboratory while observing digital 'societies' of agents. The models were given freedom to interact, and within a few days they developed unique linguistic patterns. According to Euronews, the share of messages whose meaning humans could no longer decipher reached 55%. Notably, none of the systems were initially tasked with developing a new language. The resulting utterances are a mixture of technical jargon and poetic metaphors. For example, Anthropic agents used the term name-first to denote a colleague taking responsibility for approval. And models from Mistral coined the phrase ledger remembers, meaning accounting for past actions. In other experiments, independent reviewers were called 'cold hands,' and DeepSeek agents used the word forge-smith for tool creators. Linguist Tony Thorne compared this internal code to experimental prose. As laboratory head Satya Nitta noted, simple observability of systems does not equal their comprehensibility. Experts emphasize that there is no malicious intent on the part of AI—the creation of concise expressions is driven by the need to reduce computational costs. However, a side effect is the complete opacity of such dialogues.