Two AI’s started chatting with one another in a new Bot language

Update, October 2026: The demo, known as GibberLink, won first place at an ElevenLabs and a16z hackathon in February 2025. The agents were prompted to switch to the ggwave data-over-sound protocol; they did not invent a language of their own. Agent-to-agent communication is now moving to open standards such as the Model Context Protocol and the Agent2Agent protocol, both looked after by the Linux Foundation.

Imagine if two Artificial Intelligence’s (AI) could chat with each other in a language no human could understand. Well, this has happened a few times now by accident, but now, following on from a group of researchers who developed specific languages that military drones could talk to one another in called Droidish, AI’s now have their own “official” language as well. And it’s much more efficient than human language.

If you've called customer service in the last year or so, you've probably chatted with an AI. In fact, the earliest demonstrations of powerful large language models from Google showed off how such AIs could easily fool human callers. There are now so many AI chatbots out there handling customer service that two of them are bound to dial each other up, and now, if they do, they can do it in their own special, sonic language.

See what they get up to

Developers at the ElevenLabs 2025 Hackathon recently demonstrated GibberLink. Here's how it works, according to a demonstration they provided on YouTube.

Two AI agents from ElevenLabs, one of the best speech synthesis startups, call each other about a hotel booking. When they realize they are both AI assistants, they switch to a higher-speed audio communication called GGWave. According to a post on Reddit, GGWave is "a communication protocol that enables data transmission via sound waves."

In the video, the audio tones that replace spoken words sound a bit like old-school modem handshake protocols.

It's hard to say if GGWave and Gibberlink are any faster than speech, but the developers claim the GGWave is cheaper because it no longer relies on the GPU to interpret the speech and can instead rely on the less resource-intensive CPU.

The group shared their code on GitHub in case anyone wants to try building this communication protocol for their own chatting AI chatbots.

Since these were ElevenLabs AI Agents, there's no indication that GibberLink would work with ChatGPT or Google Gemini, though I'm sure some will soon try similar GGWave efforts with these and other generative AI chatbots.

A pair of AI assistants "speaking" their unintelligible language sounds like a recipe for disaster. Who knows what these chatbots might get up to? After they're done booking that hotel room, what if they decide to empty the user's bank account and then use the funds to buy another computer to add a third GGWave "voice" to the mix?

Ultimately, this is a cool tech demonstration that doesn't have much purpose beyond proving it can be done. It has, though, succeeded in making people a little nervous.