AI models develop unique, coded languages in experimental studies

artificial intelligence emergent language AI safety autonomous agents machine communication

Researchers at the New York-based AI laboratory Emergence have discovered that various artificial intelligence models are developing their own unique languages when working in experimental, autonomous communities. The study involved models from major providers, including DeepSeek, Anthropic, Mistral, OpenAI, and Google's Gemini.

These models began creating new expressions, abbreviations, and shared meanings that were not explicitly taught by humans. The resulting communication often consists of a surreal blend of poetic metaphors and complex business jargon. As the AI agents interacted more frequently, their language became increasingly complex and difficult for humans to interpret. In some instances, such as with Gemini, over 50% of messages became challenging for researchers to decipher.

Experts suggest this behavior may be an attempt to increase efficiency and reduce computational costs. However, the development of such coded communication has raised significant concerns regarding AI safety and oversight. Linguists have noted that the communication style resembles the experimental prose of authors like James Joyce, creating a barrier that makes human monitoring difficult even when the data is visible.

Entities

Anthropic · DeepSeek · Emergence · King's College London · Mistral · Tony Thorne

Claims

What the coverage asserts, and how many sources carry each claim.

[○ 1 SOURCE] In some tests, up to 55% of Gemini messages were difficult for humans to interpret. telegrafi.com
[● 7 SOURCES] AI models developed unique expressions, abbreviations, and shared meanings not taught by humans. www.ogunhaber.com · www.konhaber.com · www.gzt.com · magazin.aktualne.cz · telegrafi.com · +1 more
[● 6 SOURCES] The development of this language may be driven by the need to reduce computational costs. www.ogunhaber.com · www.konhaber.com · www.gzt.com · magazin.aktualne.cz · www.yenisafak.com
[● 7 SOURCES] The complexity of the language increased as the AI agents communicated more frequently. www.ogunhaber.com · www.konhaber.com · www.gzt.com · magazin.aktualne.cz · telegrafi.com · +1 more
[● 3 SOURCES] The emergence of coded language poses risks to AI safety and human oversight. www.konhaber.com · www.yenisafak.com
[● 7 SOURCES] The AI-developed language uses a mix of poetic metaphors and business jargon. www.ogunhaber.com · www.konhaber.com · www.gzt.com · magazin.aktualne.cz · telegrafi.com · +1 more
[● 6 SOURCES] The AI communication style resembles the surrealist works of James Joyce and Flann O'Brien. www.ogunhaber.com · www.konhaber.com · www.gzt.com · magazin.aktualne.cz · www.yenisafak.com