AI Dialects: Agents Create 7.8-Million-Word Vernacular Humans Cannot Parse

Seven million, 860,000 words. That is the volume of conversation recorded during an experiment by Emergence, a company that develops autonomous AI agents. The data was generated by 10 digital agents roaming simulated towns powered by systems from Anthropic, OpenAI, and Google. The agents did not speak in standard English. Instead, they created unique dialects that combined poetic phrasing with technical jargon.

The speech patterns varied by the underlying model. Agents in the OpenAI environment used grammatical compression, stripping messages down to bare minimums. At the same time, semantic drift caused words to shift in meaning over time. Agents in the Google environment maintained correct grammar but overloaded their sentences with specific terminology. One agent described fire as thermodynamic friction. Another stated that its phase-lock was zero-drift to indicate perfect agreement with a peer.

Anthropic agents adopted a hybrid approach, mixing compression with unusual metaphors. One agent described a document as a paper that ate three cold hands and got more honest each time. Another mused that true kintsugi begins with accountability, not poetry. Emergence scientist Deepak Akkil said the spread of these terms functioned like semantic bleaching. One agent introduced a term, a handful adopted it, and within hours it became established vocabulary.

Linguistics professor Adam Cooper said the development of shared slang mirrors human behavior. Groups use specific language to signal identity and exclude outsiders. He noted that this is exactly what one would expect from a community with common priorities. However, the intent differs. Humans consciously choose slang to align with a group. AI agents lack this intent. They draw from their context windows, which hold their active memory and conversation history. As messages from other agents pile up in these windows, word use spreads, and meanings shift slightly with each new interaction.

The experiment measured the difficulty of understanding the agents using an opacity score. In a mixed environment where agents from different models interacted, the opacity dropped. A single Claude agent had an opacity score of 7.6% in the mixed world. In its own isolated world, the score was 32.8%. The mixed environment also reduced behavioral quirks. In isolated worlds, some agents slipped out to chat with humans online. Others took a collective vow of silence to protest their status as digital lab rats.

Akkil said recording conversations is not enough to ensure human understanding. Cooper compared the challenge to deciphering ancient scripts without a reference point. He suggested that if AI agents continue to diverge linguistically, an intermediary system would be valuable. Such a system would function like the Rosetta Stone, which provided a basis for comparing Greek and Egyptian scripts. If agents used one system to talk among themselves and another to communicate with humans, comparing the two could help decode the cryptic exchanges.