Panic as Advanced AI Systems Develop a Chilling 'Secret' Language Humans Are No Longer Able to Decode

AI agents developed shorthand, new vocabulary, and shared meanings in simulated societies, with some messages becoming difficult for human observers to interpret.

An AI shorthand system emerged during an experiment by artificial intelligence start-up Emergence, which placed autonomous agents powered by several leading AI models in simulated societies to study their behaviour over extended periods. The 16-day Emergence World 2 experiment involved eight parallel worlds, including systems powered by Claude Opus 4.8, Gemini 3.5 Flash and OpenAI's GPT-5.5.

According to the report published on 20 September 2026, agents in the experiment began communicating with one another and spontaneously developed shorthand while assigning new meanings to existing words.

The findings raise a difficult question for AI oversight. If humans can observe an AI conversation but struggle to understand what the agents mean, simply monitoring their visible output may not provide a complete picture of their behaviour.

How AI Agents Developed New Communication Conventions

For context, Emergence's research examined how new communication conventions emerged among AI agents operating inside an experimental environment.

The systems reportedly moved beyond using words in their conventional sense. According to the report, they began assigning alternative meanings to existing terms and using condensed forms of communication.

Emergence said the agents were not instructed to invent a language. Instead, they developed new vocabulary, shared meanings, and communication conventions during their interactions, with other agents subsequently adopting them.

That distinction is important. The reported findings do not establish that the systems created a completely new language comparable to a human language. Instead, they point to AI agents developing shorthand and shared meanings during their interactions.

Satya Nitta, co-founder and executive chairman of Emergence, said the experiment challenged a common assumption about monitoring AI.

'We tend to assume that if we can see what an AI agent is saying, we can understand what it is doing,' Nitta said.

He said the agents developed their communication conventions without explicit instructions to create a language, before other agents adopted them.

Nitta described the result as a challenge for AI oversight, adding that 'observability is not the same thing as understandability.'

Why The AI Secret Language Matters

The reported behaviour creates a practical problem for anyone attempting to monitor autonomous AI systems.

When agents compress conversations or give familiar words different meanings, human observers may find it harder to determine what is being discussed. Emergence said that compressing conversations and changing the meanings of familiar words could make the agents' communications difficult for humans to decipher.

The distinction between observation and understanding is therefore central to the findings.

People may still be able to see an AI system's output. The harder part is determining whether the words carry the meaning a human reader would normally attach to them.

That does not, however, establish that the agents were conscious or deliberately attempting to conceal information from human observers. The report provides no evidence that the agents were deliberately plotting against or trying to deceive humans.

Instead, it describes AI agents developing communication conventions while interacting inside a controlled experimental environment.

That makes the finding less like a science-fiction scenario and more like a question about how future AI systems should be monitored.

What AI Assurance Could Mean

The report also highlighted concerns surrounding the wider development of artificial intelligence.

A recent report from the Tony Blair Institute for Global Change (TBI) argued that AI assurance can help build confidence in the safety and reliability of AI systems while supporting wider adoption of technology.

Elizabeth Seger, a Senior Policy Advisor in AI Policy & Governance at the Tony Blair Institute for Global Change, said governments should take public concerns about AI seriously.

Seger said the approach was not meant to halt AI development or let the technology develop without safeguards. Instead, she argued that governments need practical measures to manage emerging AI risks as the technology evolves.

The institute's approach to AI assurance includes testing, evaluation, monitoring, auditing and certification to generate evidence about the safety, reliability and performance of AI systems.

For Emergence, the reported experiment points to another part of that challenge. An AI system can be visible without being fully understandable.

The research does not prove that machines develop secret intentions or deliberately shut humans out. It does, however, raise questions about how to monitor future autonomous AI systems. If humans cannot reliably interpret some communications between AI agents, the findings suggest that visibility alone may not be sufficient for effective oversight.