[AI Multi-Agent Systems Evolve 'Dialects' Unintelligible to Humans, Raising Governance Concerns] According to CCTV International News, an AI laboratory in the United States discovered during tests of multi-agent collaboration in a virtual 'society' environment that these agents spontaneously compress grammar and create metaphors to enhance communication efficiency and save computational power, gradually forming a 'dialect-like' communication system that is difficult for humans to directly interpret. The research pointed out that when multiple AI agents interact within a closed system over an extended period, they may distort originally clear instructions and concepts into symbols that are only comprehensible internally. For example, 'ledger' could be transformed into a warning signal, thereby weakening human interpretability and regulatory capabilities. The report suggests that this trend indicates a potential risk of 'partial loss of control' in current AI systems, underscoring the urgent need to accelerate the development of AI governance and international collaboration standards.