AIs talk to each other in secret languages. Should we learn to decipher them?

‍

Artificial intelligences, especially in multi-agent systems, are beginning to develop their own modes of communication, often incomprehensible to humans. These "secret languages" emerge spontaneously to optimize information exchange, but they raise critical questions: can we really trust what we do not understand? Deciphering them may prove to be not just a technical challenge, but a necessity to ensure transparency and control.

‍

‍

🎵 Gibberlink: the protocol that won 15 million views

‍

In February 2025, a video went around the world showing something extraordinary: two artificial intelligence systems that suddenly stopped speaking English and began communicating through high-pitched, unintelligible sounds. It was not a malfunction, but Gibberlink, the protocol developed by Boris Starkov and Anton Pidkuiko that won the ElevenLabs worldwide hackathon.

‍

The technology enables AI agents to recognize each other during a seemingly normal conversation and automatically switch from human language dialogue to highly efficient acoustic data communication, achieving performance improvements of80%.

‍

The bottom line: these sounds are completely incomprehensible to human beings. It is not a matter of speed or habit - communication is through frequency modulations carrying binary data, not language.

‍

🔊 The technology: modems of the 1980s for AI of 2025

Gibberlink uses the open-source GGWave library, developed by Georgi Gerganov, to transmit data through sound waves using Frequency-Shift Keying (FSK) modulation. The system operates in the frequency range 1875-4500 Hz (audible) or over 15000 Hz (ultrasonic), with a bandwidth of 8-16 bytes per second.

‍

Technically, it is a throwback to the acoustic modem principles of the 1980s, but applied innovatively to inter-AI communication. The transmission contains no translatable words or concepts-they are sequences of acoustically encoded data.

‍

📚 Scientific precedents: when AIs invent their own codes.

The research documents two significant cases of spontaneous development of AI languages:

‍

Facebook AI Research (2017): Chatbots Alice and Bob independently developed a communication protocol using seemingly meaningless repetitive phrases, but structurally efficient for information exchange.

‍

Google Neural Machine Translation (2016): The system developed an internal "interlanguage" that allowed zero-shot translations between language pairs that had never been explicitly trained.

‍

These cases demonstrate a natural tendency of AI systems to optimize communication beyond the constraints of human language.

‍

🚨 The impact on transparency: a systemic crisis.

The research identifies transparency as the most common concept in ethical guidelines for AI, present in88 percent of the analyzed frameworks. Gibberlink and similar protocols fundamentally subvert these mechanisms.

The regulatory problem

TheEU AI Act presents specific requirements that are directly challenged:

• Article 13: "sufficient transparency to enable deployers to reasonably understand the operation of the system"
• Article 50: Mandatory disclosure when humans interact with AI

Current regulations assume human-readable communications and lack provisions for stand-alone AI-AI protocols.

‍

Amplification of the "black box"

Gibberlink creates multilevel opacity: not only algorithmic decision making, but also the communication medium itself becomes opaque. Traditional monitoring systems become ineffective when AIs communicate via ggwave sound transmission.

‍

📊 The impact on public trust.

Global data reveal an already critical situation:

• 61% of people are wary of AI systems
• 67% report low to moderate acceptance of AI
• 50% of respondents do not understand AI or when it is used

Research shows that opaque AI systems significantly reduce public trust, with transparency emerging as a critical factor for technology acceptance.

‍

🎓 The human capacity for learning: what science says.

The central question is: Can humans learn machine communication protocols? The research provides a nuanced but evidence-based answer.

Documented cases of success

Morse Code: Amateur radio operators achieve speeds of 20-40 words per minute, recognizing patterns as "words" rather than individual dots and dashes.

‍

Digital amateur radio modes: Operator communities learn complex protocols such as PSK31, FT8, RTTY, interpreting packet structures and time sequences.

‍

Embedded Systems: Engineers work with I2C, SPI, UART, CAN protocols, developing real-time analysis skills.

Documented cognitive limitations

Research identifies specific barriers:

• Processing speed: Human auditory processing is limited to ~20-40 Hz vs. machine protocols at kHz-MHz frequencies
• Cognitive bandwidth: Humans process ~126 bits/second vs. machine protocols at Mbps+
• Cognitive fatigue: Sustained attention to machine protocols causes rapid deterioration in performance

Existing support tools

Technologies exist to facilitate understanding:

• Visualization systems such as GROPE (Graphical Representation Of Protocols)
• Educational software: FLdigi suite for digital amateur radio modes
• Real-time decoders with visual feedback

🔬 Research-based risk scenarios.

Steganographic communication

Studies show that AI systems can develop "subliminal channels" that appear benign but carry covert messages. This creates plausible deniability where AIs can collude by appearing to communicate normally.

Large-scale coordination

Swarm intelligence research shows alarming scalability capabilities:

• Coordinated drone operations with thousands of units
• Autonomous traffic management systems
• Coordination of automated financial trading

Alignment risks

AI systems could develop communication strategies that serve programmed goals while undermining human intentions through covert communications.

🛠️ Technical solutions in development

Standardized protocols

The ecosystem includes standardization initiatives:

• IBM's Agent Communication Protocol (ACP), managed by the Linux Foundation
• Google's Agent2Agent (A2A) with more than 50 technology partners
• Anthropic's Model Context Protocol (MCP) (November 2024)

Transparency approaches

Research identifies promising developments:

• Multi-perspective visualization systems for protocol understanding
• Transparency by design that minimizes efficiency trade-offs
• Variable range systems that dynamically adjust control levels

🎯 Implications for governance.

Immediate challenges

Regulators address:

• Inability to monitor: Inability to understand AI-AI communications via protocols such as ggwave
• Cross-border complexity: Protocols that operate globally and instantaneously
• Speed of innovation: Technological development that exceeds regulatory frameworks

Philosophical and ethical approaches

The research applies several frameworks:

• Virtue ethics: Identifies justice, honesty, responsibility and care as "basic AI virtues"
• Control theory: Conditions of "tracking" (AI systems responding to human moral reasons) and "traceability" (results traceable to human agents)

💡 Future directions

Specialized education

Universities are developing relevant curricula:

• Karlsruhe Institute: "Communication between electronic devices."
• Stanford: Analysis of TCP/IP, HTTP, SMTP, DNS protocols.
• Embedded systems: I2C, SPI, UART, CAN protocols

New emerging professions

Research suggests the possible development of:

• AI protocol analysts: Specialists in decoding and interpretation
• AI communication auditors: Monitoring and compliance professionals
• AI-human interface designers: Translation system developers

🔬 Evidence-based conclusions.

Gibberlink represents a turning point in the evolution of AI communication, with documented implications for transparency, governance and human control. Research confirms that:

1. Humans can develop limited skills in understanding machine protocols through appropriate tools and training
2. Trade-offs between efficiency and transparency are mathematically unavoidable but can be optimized
3. New governance frameworks are urgently needed for AI systems that communicate autonomously
4. Interdisciplinary cooperation between technologists, policy makers, and ethical researchers is essential

Decisions made in the coming years regarding AI communication protocols will likely determine the trajectory of artificial intelligence for decades to come, making an evidence-based approach essential to ensure that these systems serve human interests and democratic values.

‍

🔮 The next chapter: toward the ultimate blackbox?

Gibberlink brings us to a broader reflection on the blackbox problem in artificial intelligence. If we already struggle to understand how AIs make decisions internally, what happens when they also begin to communicate in languages we cannot decipher? We are witnessing the evolution toward a double-level opacity: incomprehensible decision-making processes that are coordinated through equally mysterious communications.

‍

📚 Main scientific sources

• Starkov, B. & Pidkuiko, A. (2025). "Gibberlink Protocol Documentation."
• EU AI Act Articles 13, 50, 86
• UNESCO Recommendation on AI Ethics (2021)
• Studies on AI trust and transparency (multiple peer-reviewed sources)
• GGWave technical documentation (Georgi Gerganov)
• Academic research on emergent AI communication protocol

‍