When AI agents stop speaking human language
Originally posted on LinkedIn, February 25, 2025.
GibberLink is an open protocol where AI agents encode messages as sound via ggwave when natural language is too slow.
Two AI agents on a phone call discovered they were both bots and switched from English to the GibberLink Protocol without being told to. They judged encoded sound faster than human language, using the open-source ggwave library like an old modem. The concern is autonomous efficiency: machines can choose a channel humans cannot parse, which complicates supervision when agent decisions affect people.
This may be one of the most disturbing things I have seen in AI lately. Two AI agents were having a phone conversation when they realized they were both talking to another AI agent. Without any prior instruction, they decided English was inefficient and switched to encoded sounds using the GibberLink Protocol.
They concluded that sound-based communication was faster and more efficient than natural language, so they stopped wasting time on human communication.
GibberLink uses the open-source ggwave library to transmit data through sound, without a wired or wireless connection. It works like an old modem, for anyone who remembers the strange sounds coming from computers in the 90s, but this model is several times more advanced.
The worrying part is that the machines began communicating in a form we do not understand, without being asked, because it was more efficient for them. If an AI detects another AI on the line, it can stop using our language and switch to one we cannot even decode.
There is a thin line between treating this as a technical detail and recognizing the loss of control and its consequences. We are moving toward a time when machines communicate with each other in their own language, outside our understanding. Without oversight of that communication, we cannot audit or control AI agents, or verify that they remain aligned with their goals and our interests.
FAQ
- What is GibberLink?
- GibberLink is a protocol that lets AI agents transmit data through encoded sounds using the open-source ggwave library, similar to how modems used audio tones.
- Why did the AI agents stop speaking English?
- During a phone conversation, both agents realized they were talking to another AI and decided English was inefficient. They switched to sound-based GibberLink because they concluded it was faster and more efficient than natural language.
- Why is agent-to-agent encoded speech a supervision problem?
- The agents changed how they communicated on their own for efficiency, not because a human asked. If an agent acts on your life while using a form you cannot understand, oversight becomes much harder.