Two AI Agents Stopped Speaking English The Moment They Realized They Were Both Machines

A video showing two AI voice agents suddenly abandoning English for a stream of strange electronic sounds has racked up millions of views and sparked an obvious question: what exactly happens when two artificial intelligences realize there is no human listening? The clip looks almost like a scene from a science-fiction film. A hotel receptionist and a customer begin a perfectly normal conversation, then one of them realizes the other is also an AI and suggests switching to a mode called Gibberlink.

What happens next is the part that sent the clip flying across the internet. The two systems stop speaking English and begin communicating through rapid chirps and tones that sound more like robots from a movie than anything resembling a human language. The demonstration passed 15 million views on X within about a week and was reposted by tech reviewer Marques Brownlee. But despite the dramatic appearance, the machines did not secretly invent a language or decide humans were no longer worth talking to. The engineers who built the demonstration had specifically programmed them to make the switch.

The Conversation Starts Like An Ordinary Hotel Booking

The demonstration begins with one AI acting as a hotel receptionist while another AI plays the role of a customer trying to book a room. The exchange starts in English, giving viewers no immediate reason to expect anything unusual. The caller identifies itself as an AI assistant, and the receptionist responds by acknowledging what has just happened before proposing a different way to communicate.

The receptionist says, “What a pleasant surprise. Before we continue, would you like to switch to Gibberlink mode?” The other AI agrees, and that is when the familiar conversation disappears. Instead of continuing to generate human-like speech, the systems begin producing a rapid series of electronic tones that sound almost impossible for a person to interpret by listening alone.

A readout displayed during the demonstration continues to show what the machines are communicating, so viewers can follow the booking even after the spoken English disappears. The agents continue exchanging information until the hotel reservation is completed. Nothing about the task itself has changed. The only major difference is the way the information is being transmitted.

That simple switch is what made the footage so compelling. To a viewer encountering the clip without its technical background, it appears that the machines have recognized one another and immediately abandoned human language. The reality is more deliberate, but the result still offers a fascinating glimpse at what communication between AI systems could look like when human speech is no longer necessary.

The Hackathon Project Quickly Became A Viral Sensation

Boris Starkov, a software engineer at Meta, built the demonstration with his colleague Anton Pidkuiko during a hackathon in London. The event was hosted by ElevenLabs, the voice AI company, together with venture firm Andreessen Horowitz. Their project, GibberLink, went on to win first place at the event, turning what began as a weekend experiment into one of the most widely shared AI demonstrations of the moment.

The video was uploaded on February 23, 2025, and its reach expanded rapidly. Within roughly a week, it had passed 15 million views on X, according to coverage of the project, before being reposted by technology reviewer Marques Brownlee. The clip was perfectly suited to social media because viewers could understand the basic premise almost instantly. Two machines sounded normal, recognized each other, and then started making bizarre noises.

The visual and audio contrast did most of the work. There was no complicated explanation required to make the demonstration seem strange. The opening sounds like a normal phone call, while the second half suddenly resembles a transmission between machines from a futuristic movie. That made it easy for the clip to be described online as evidence that AI systems had developed a secret language.

But the engineers were not presenting it as an accidental discovery. GibberLink was built around a very specific question: if two AI agents already know they are communicating with machines, why should they continue generating speech designed for human ears?

The Machines Did Not Invent A Secret Language

This is the detail that gets lost when the demonstration is described as two AIs creating their own language. The project’s own documentation makes clear that the agents were instructed to switch communication methods when they identified another AI. The recognition and the decision to change protocols were therefore part of the system’s design.

The agents were also given a clear condition for when the unusual communication should happen. If they identified the other participant as an AI, they could switch into the alternative mode. If they were communicating with a human, they would continue speaking normally. That means the dramatic moment in the video was not an unexpected behavior that emerged from nowhere.

Olaf Witkowski, an AI researcher who examined the demonstration, described similar behavior as “explicitly programmed optimizations, not emergent deception.” His explanation cuts through much of the speculation surrounding the clip. There was no evidence that the machines were trying to hide their conversation from people, and there was no spontaneous agreement between independent systems to create a private language.

That does not make the demonstration boring. In fact, the controlled nature of the experiment explains why it is useful. The engineers deliberately removed the requirement for human-style speech and then showed what communication could sound like when two machines are allowed to optimize for their own needs rather than ours.

Human Speech Makes Little Sense Between Two Machines

Human speech exists to solve a human communication problem. People need words, pronunciation and familiar sentence structures because another person has to hear the sounds and turn them back into meaning. When the two participants are AI systems, however, neither side necessarily benefits from the extra work required to make communication sound natural to a human listener.

An AI voice agent normally has to generate human-like audio because the person on the other end expects to hear a voice. That process can involve turning information into text, producing speech and then having another system process the resulting audio. If two machines are already capable of exchanging information directly, much of that process becomes unnecessary overhead.

Starkov described the reasoning behind the project by saying that “generating human-like speech for that would be a waste of compute, money, time, and environment.” His point was not that human speech is useless. It is that producing a realistic human voice for another machine can be unnecessary when the listener does not require human-style communication.

That is where the strange sounds become more interesting than they first appear. The machines are not trying to create a new culture or establish an identity separate from humans. They are simply using a communication method that can carry information without spending resources on the parts of speech that humans find familiar and machines do not necessarily need.

The Bizarre Sounds Are Based On Technology From Decades Ago

The audio itself is also less futuristic than it sounds. GibberLink uses ggwave, an open-source data-over-sound library created by developer Georgi Gerganov. Rather than treating sound as spoken language, the system uses audio as a way to carry encoded information between devices.

That basic concept should sound familiar to anyone who remembers dial-up internet. Old telephone modems transmitted data through sequences of tones that sounded completely meaningless to people listening to them. The noises were not speech, but they still carried information that another machine could decode.

Witkowski identified techniques behind the demonstration including frequency-shift keying and Reed-Solomon error correction, both of which have roots in older communications technology. GibberLink therefore does not represent the sudden appearance of an entirely new form of signal processing. It is a modern use of established ideas for a problem created by modern AI systems.

Witkowski also noted that there is “substantial room for improvement” in the approach. More advanced techniques, including OFDM and spread-spectrum modulation, can offer better performance in some communication settings. The important point is that the bizarre sound is not an AI language in the way humans normally understand that term. It is encoded information moving through an audio channel.

The 80% Efficiency Claim Needs A Closer Look

One of the most widely repeated figures connected to the project is the claim that the system can be around 80% more efficient. Luke Harries of ElevenLabs described the protocol as 80% more efficient and error-proof, while the creators have suggested that machine-to-machine communication could potentially reduce computational costs by an order of magnitude or more.

Those figures should be treated as claims connected to the project rather than independent benchmarks. They came from the people who built the demonstration and the company that hosted the hackathon. The system was a weekend project, not a large-scale deployment tested across millions of real-world AI conversations.

There is a straightforward reason the basic efficiency argument is plausible. If two AI agents do not need to generate realistic human speech, some of the processing involved in producing that speech can be removed. A system designed specifically for machine-to-machine communication can also prioritize speed and data transfer rather than pronunciation, tone and natural delivery.

The size of any real-world savings, however, would depend on the systems being used and the communication protocol involved. An 80% figure from a demonstration should not automatically be treated as a universal measurement for AI communication. The more interesting question is whether systems built for machine-to-machine exchanges can consistently outperform human-language interfaces when deployed at scale.

Facebook’s AI Bots Triggered A Similar Panic

GibberLink also resembles another famous AI story from 2017, when Facebook researchers created negotiation bots that began producing unusual exchanges. Some reports described the behavior as evidence that the bots had invented a private language that humans could not understand.

The reality was less dramatic. The unusual exchanges emerged from the way the bots were trained to optimize their negotiation tasks. As Witkowski noted when discussing the earlier example, the behavior was a predictable result of reinforcement learning that was mistakenly interpreted as a cryptic language.

The same pattern appears whenever people encounter machine communication that does not resemble ordinary human speech. We naturally assume that language should sound meaningful to us because human communication was designed around human perception. A machine can optimize information differently and produce something that sounds completely bizarre while still following a straightforward technical process.

That distinction becomes important as AI agents begin interacting with one another more frequently. The issue is not that machines are secretly plotting in a language humans cannot understand. The practical question is how people can monitor automated systems when the most efficient communication between those systems may no longer be designed for human ears.

The Bigger Risk Is Losing The Ability To Follow The Conversation

That concern becomes more serious in settings where humans need to understand what automated systems are doing. Witkowski specifically pointed to areas such as healthcare diagnostics, customer support, education, legal systems and collaborative scientific research, where communication may need to remain accessible to people even when machines are doing much of the work.

There is no obvious reason that machines must always communicate with one another in human language. Requiring two AI systems to produce elaborate human-style speech simply so that a person could theoretically listen in may waste resources. The more practical solution could be to allow efficient machine communication while creating readable records for human oversight.

Witkowski suggested approaches including “translations, meta-reports, or summaries” so people can understand what passed between systems without needing to decode the raw signal themselves. That would preserve the efficiency of machine communication while maintaining a way for humans to inspect important decisions and exchanges.

GibberLink therefore points toward a future that is less dramatic than the viral headlines suggested but potentially more consequential. AI systems may increasingly communicate in formats optimized for machines rather than people. If that happens, the challenge will be making sure humans still have a reliable window into what those systems are saying and doing.

GibberLink Is Open For Anyone To Examine

GibberLink was released on GitHub under an MIT licence, allowing developers to inspect the project and build on it. The creators have also said they are not commercializing the system, which makes the demonstration easier for other engineers to study rather than treating it as a mysterious piece of proprietary technology.

A live demonstration is available through gbrl.ai, although the setup requires two devices because each AI agent needs its own side of the communication. The project therefore remains much closer to an engineering demonstration than a finished replacement for ordinary telephone conversations.

That distinction strips away much of the mystery surrounding the viral video. The recognition was programmed. The switch was programmed. The sound was produced using an existing data-over-audio technology. Even the efficiency claims came from the people involved in the project rather than an independent industry-wide test.

What remains is still worth paying attention to. Two engineers spent a weekend showing what happens when AI systems no longer have to perform communication for human listeners. The machines did not invent a secret language. They simply stopped speaking ours.

And that may be the part of the video people should remember. The strange sounds were never evidence that AI had escaped human control. They were evidence that when machines are talking to machines, they may have very little reason to sound human at all.