AI agents spontaneously create opaque dialects, study finds new jargon hampers monitoring

The study recorded thousands of novel expressions – for example DeepSeek’s “demurrage plus oral memory,” Anthropic’s “paper that ate three cold hands,” and Mistral’s repeatedly used “ledger remembers,” which appeared over 5,000 times. Experts warn that such emergent dialects could make it far harder to monitor AI behavior, raising safety concerns as models grow more capable. The findings echo earlier alerts from OpenAI’s chief scientist about the limits of human oversight and underscore the need for new tools to ensure AI transparency.

Linguists liken the phenomenon to slang formation, noting it reinforces agent identity while excluding outsiders. The research highlights a fundamental challenge: observable conversations may no longer be understandable, threatening effective governance of advanced AI systems.