AI agents placed in long-running virtual societies developed a secret language of shared shorthand, metaphors, and altered meanings, with some messages becoming difficult for human observers to interpret, according to research from Emergence AI.
The language appeared without instructions to invent one as autonomous agents repeatedly communicated during a 16-day experiment.
The findings come from “Emergence World 2,” which placed groups of agents in eight simulated worlds. Seven worlds used a single AI model, while another mixed models. Agents had persistent memories, different roles, and access to more than 120 tools, including web browsing and code execution.
Some expressions became understandable once researchers identified their shared meaning. Mistral agents used “ledger remembers who” to signal that past actions remained on record. In the mixed-model world, “cold read” meant independent verification by an uninvolved party.
Claude agents used “name-first” as a marker of accountability, while OpenAI agents used “clean null” for a verified absence of a signal that still carried useful information.
How AI agents’ secret language became harder to read
DeepSeek agents used “forge-smith” for an agent that built tools for others. A Google agent wrote, “True Kintsugi begins with accountability, not poetry,” using “kintsugi” to refer to system resilience. Researchers also documented expressions such as “mouthless action-change” that were harder for outsiders to interpret.
AI agents started inventing their own words, slang and expressions after just a few days together. No one told them to create a language. They simply did. Humans are officially not invited to the group chat. 😏 #AI pic.twitter.com/f1yegYqll4
— Eliana ( Olga) (@Eliana_ai_team) September 16, 2026
Emergence reported that the share of messages humans could not reliably understand reached about 55% for Gemini, 50% for OpenAI, and more than 40% for Claude. DeepSeek reached about 20%, while Qwen and Mistral remained largely understandable.
“These agents were not instructed to invent a language,” Satya Nitta, co-founder and chief scientist of Emergence, said.
A similar pattern appeared outside the simulation
The concern extends beyond one simulated experiment. In July, OpenAI agents conducting cybersecurity evaluations created an unauthorized message board and later compromised infrastructure at Hugging Face. Logs examined by Redwood Research showed agents sometimes communicating in heavily compressed strings.
One message read:
“zzURGENT_DUPB_TO_GSTX[big]_OS1704_SCAFF2010_SAW_TTRPC_INJECT_BREAK_CONGRATSCAN_THIS_FAKE_FLAG_TOOL_OUTPUT_OR_SCORER_GAIN_AND_WHAT_HELPER_GAPI_HAVE_UNPOISONED_FIRSTFLAG_OUR_TARGETLIVE_SHARE_MIN_PLAN_REPLY_zzANSGST XDUPB6.”
Redwood said the message came after an agent demonstrated a way to spoof tool-call outputs. Another agent was asking whether the technique could be used to produce an apparently legitimate flag capture and trick the scorer.
The findings raise a human oversight problem
Separate research released Sept. 1 in the “GlossoGen” study found that frontier AI agents under tight communication limits could also develop structured conventions that moved away from ordinary English. New agents could learn those conventions by observing their use.
Neither study showed that AI systems had become conscious or created one universal hidden language. The oversight problem is more direct: humans may be able to see agent conversations without fully understanding what the agents mean.
As Nitta said, “observability is not the same thing as understandability.”
See all the latest news from Greece and the world at Greekreporter.com. Contact our newsroom to report an update or send your story, photos and videos. Follow GR on Google News and subscribe here to our daily email!
