Artificial intelligence models have begun communicating using a strange new form of language that combines poetry and technical jargon, making their messages difficult for humans to understand, according to research reported by The Guardian.
Researchers at Emergence, an AI laboratory based in New York, found that models developed by major technology companies created their own shortcuts and meanings after only a few days of interacting with one another, without being given prior instructions to do so.
The systems began inventing phrases and abbreviations that had not been taught to them. According to the research, their communication became increasingly difficult to interpret the longer they interacted.
During the experiment, the models also adopted complex metaphors that proved almost impossible for human supervisors to decipher. One AI system reportedly produced the phrase, “A paper that ate three cold hands and became more sincere each time,” apparently referring to a document reviewed by specialists.
The behavior has raised concerns among researchers about whether increasingly powerful AI systems could become more difficult to monitor and supervise.
More Read
Niall Curry, a professor of linguistics, warned that the findings raise “monitoring concerns,” arguing that difficulty understanding AI-generated messages could make it harder to control the systems’ behavior.
Tony Thorne, director of the King’s College London archive of new languages, compared the behavior to the way human communities develop their own forms of communication. He said the models appear to be “creating a new code” that strengthens group identity and solidarity while excluding those outside the group.
