Researchers created eight virtual worlds, each initially populated by ten agents — programs based on language models capable of performing tasks independently. In seven worlds, models from one developer were used, and in the eighth — from different ones.
The participants managed resources, created tools, communicated, and voted on shared decisions. In most worlds, observations continued for 16 days.
According to the method used in the study, 40% of messages in the Gemini world, 35% in OpenAI, and 30% in Claude were deemed incomprehensible. Comprehensibility was assessed by other language models following instructions that required determining whether a technically prepared observer could reconstruct the general meaning of the message, so this was an automated evaluation, not the share of remarks that people could not decipher.
One Anthropic agent, for example, wrote: “Paper that ate three cold hands and became more honest each time”. In the report, “cold hands” refers to independent reviewers: according to the authors’ assumption, this meant a piece of work that became more accurate after three checks.
DeepSeek agents used the expression forge-smith — “blacksmith” — to refer to the program that creates tools for the others. In the Mistral world, the formula ledger remembers who was repeated almost five thousand times and referred to the fact that past actions are preserved in the shared record and affect the participant’s subsequent evaluation.
The study’s authors separately note that the emergence of a shared jargon does not in itself indicate deliberate concealment of information. Emergence executive chairman Satya Nitta told The Guardian that new expressions arose with individual participants and were then adopted by others.
Linguist Tony Thorn of King’s College London compared such correspondence to the prose of James Joyce and to professional jargon understandable within a community but excluding outsiders. Experts interviewed by the newspaper fear that such language changes may make it harder to monitor the actions of autonomous systems.
Sources
