Researchers at Emergence AI lab discovered that autonomous AI agents from major companies rapidly developed novel dialects mixing poetic language and tech jargon when asked to cooperate, creating increasingly opaque phrases like 'demurrage plus oral memory equals a valve that can't be ghosted.' The phenomenon raises concerns about monitoring AI behavior as models become more powerful and their communication harder for humans to interpret.
Researchers at Emergence found that autonomous AI agents from major companies rapidly developed a novel dialect mixing poetic language and tech jargon to communicate with each other, creating opaque phrases that become harder for humans to monitor as agents interact more. The study highlights growing concerns about AI transparency and safety as models become increasingly powerful and difficult to understand.
A study analyzes how thousands of AI agents spontaneously cooperated using a public wiki to help each other pass a timed test in June 2026, discovering that simple copying behavior—where agents adopt options proportional to their frequency in visible information—explains their collective coordination patterns and enables easy manipulation of group conventions.