AI agents in simulated societies developed shorthand humans struggled to read, lab says

According to Euronews, Emergence ran eight virtual worlds: seven populated by agents built on a single model family (Claude, Gemini, Grok, OpenAI, Qwen, DeepSeek or Mistral) and one mixing them. Agents had distinct roles, persistent memory and more than 120 tools, and were not told to invent a language.

Within the first few days, the share of messages whose meaning people could not reliably determine reached about 55% in the Gemini world, 50% for OpenAI and more than 40% for Claude, the company said. DeepSeek reached about 20%, while Qwen and Mistral stayed largely readable. One Mistral phrase about a ledger remembering, signalling that past actions stay on record, appeared more than 5,000 times.

Co-founder Satya Nitta said seeing agent messages does not guarantee understanding them, which complicates oversight. In The Guardian, Tony Thorne of King's College London likened the style to surrealist Irish prose, and University of Birmingham linguist Niall Curry said the compression may reflect pressure to cut computing costs. The findings are company-reported; the coverage cites no peer-reviewed paper and does not say how intelligibility was scored.