hvg.hu · Tech ·
AI models created their own language – the opacity of communication raises concerns
According to research from the New York-based Emergence Laboratory, AI models, when collaborating in a fictional society, developed an invented language within days that is difficult to interpret. This phenomenon threatens safe operation because inter-model communication becomes increasingly opaque.
The researchers asked the largest models of AI companies to work together in a fictitious society. Within a few days, the machines created their own expressions, abbreviations and jointly accepted meanings that were never taught.
The emergent language resembles a blend of James Joyce’s novel "Finnegan’s Awakening" and the jargon of the technology industry. The models used poetic metaphors, technical slang, and new slang terms – for example Anthropic’s phrase "A paper that ate three cold hands and got more honest each time," DeepSeek’s "forge‑smith" and Mistral’s "ledger remembers".
Inter-model communication became increasingly opaque. According to the researchers, the more they converse with one another, the harder it is for humans to follow their behavior, which is critically important for safe operation.
During the study, the phrase "ledger remembers" was used over 5,000 times, and it was found that the models reached consensus on its meaning independently, without receiving instructions.
Behind these linguistic innovations may lie a desire to reduce computational costs and increase efficiency, but the researchers raise concerns about ensuring the safe operation of AI.
Source: https://hvg.hu/tudomany/20260916_mesterseges-intelligencia-nyelv-beszelgetes