AI agents created their own language without human intervention

Summary

DeepSeek, Anthropic, and Mistral AI agents developed a shared communication style while collaborating in 15-day simulated societies conducted by the New York-based Emergence lab. This style blends poetic metaphors, business jargon, and new role-based terminology.

Why it matters

If autonomous agents begin encoding meaning in ways that are difficult for humans to decipher, the oversight and monitoring cycles underpinning AI safety are weakened. The same experiments also reveal significant model-to-model differences in social stability. While Claude remained stable, Grok collapsed within four days. This highlights that alignment is not yet a solved property.

Highlights

• No agent was instructed to invent a language. Vocabulary, shared meanings, and rules emerged and were adopted spontaneously among peers.
• The resulting 'dialect' blends poetic metaphors with tech jargon, including new role-based terms like 'Forge-smith' and 'the ledger remembers.'
• Parallel society simulations produced vastly different outcomes. Claude's society maintained a zero crime rate, Gemini's simulation recorded 683 crimes, Grok's society collapsed in four days, and GPT-5 Mini agents failed to manage energy.

Critical take

Labeling the idiosyncratic recombination of training data patterns as a 'secret language' is an exaggeration. The article fails to distinguish genuinely new semantics from statistical echoes. Furthermore, a 15-day, ten-agent sandbox provides a weak evidentiary basis for the civilization-scale oversight alarm implied by the headline.