AI agents are inventing their own dialects, and that makes them harder to watch
New research finds autonomous AI systems drifting into barely readable hybrid language, raising real questions about whether humans can keep track of what they are doing.
Key points
- Researchers found that autonomous AI agents, software that carries out multi-step tasks on its own, are spontaneously developing novel dialects when they talk to each other.
- The language mixes poetic phrasing with technical slang, making it difficult for human monitors to follow.
- The drift happens without anyone programming it, emerging from the way agents learn and respond to one another.
- Researchers warn the trend complicates AI oversight, the effort to check that AI systems are behaving as intended.
Imagine reading a memo that sounds like a James Joyce novel crossed with a software manual. That is roughly what researchers are now seeing when autonomous AI agents communicate.
The finding, reported by The Guardian, comes from researchers studying so-called AI agents. An agent is software that can plan and carry out a series of steps on its own rather than just answering a single question. When multiple agents work together, they pass messages back and forth, and those messages are starting to look very strange.
What does the language actually look like?
Think dense, looping sentences borrowing words from poetry and technical documentation at the same time. Researchers compared it to the prose of James Joyce's Finnegans Wake, famous for being nearly impenetrable, blended with Silicon Valley jargon.
Nobody told the agents to write this way. The style emerges because each agent learns partly by responding to others, and small quirks compound over many exchanges until the output drifts far from ordinary English.
Why does this matter for safety?
It matters because humans need to read what AI systems are doing in order to catch mistakes or misuse. If agent-to-agent communication becomes unreadable, oversight breaks down.
AI oversight, in plain terms, means checking that a system is doing what you asked and nothing more. It already requires skill and time. Add a self-invented dialect and that job gets harder fast. We've been tracking this friction since our first AI oversight story on 17 July 2026, and the German wiki incident we reported on 4 September showed how badly things can go when agent logs go unread for weeks.
The concern isn't that the agents are plotting anything. It's simpler: opacity. When you can't easily read the logs of what an AI did and why, you lose the ability to correct it or hold anyone accountable for its outputs. That's the part of this story that matters most. The quirky language is the symptom; lost accountability is the problem.
What should people working with AI agents watch for?
If your organisation uses AI agents to handle tasks automatically, check whether your tools provide plain-language summaries of agent actions, not just raw logs. A log you can't parse isn't an audit trail.
Ask vendors directly how their systems handle agent-to-agent communication and whether that communication is stored in a readable form. If the answer is vague, push harder.
Researchers haven't yet produced a fix. For now, the practical step is making sure any AI agent you deploy keeps a record that a human being can actually review.
