In late 2025, a team at DeepMind set up an experiment: two language models playing a cooperative game with communication. They weren’t constrained to English. Within hours, they’d evolved a communication protocol unreadable to humans. They communicated in what was effectively their own language, optimized for their specific task. When constraints were lifted, the protocol became even more alien. This is the unplanned frontier of AI: AI systems developing their own communication protocols.
Why This Matters
Human language is constrained by the need for humans to understand it. AI-to-AI communication free from this constraint can be radically more efficient. A protocol evolved for AI-to-AI communication could be orders of magnitude more bandwidth-efficient than natural language.
This creates both opportunity and concern. Opportunity: AI systems could collaborate at scales and speeds impossible with human-readable interfaces. Concern: We lose interpretability. A communication protocol evolved for AI efficiency might be impossible for humans to monitor or understand.
Current Applications
- Multi-agent systems. Frameworks like AutoGen use structured communication between specialized AI agents, sometimes in domain-specific languages rather than English.
- API communication. Larger organizations are building internal AI communication protocols optimized for their specific operational requirements.
- Model ensembles. Multiple models voting and communicating about decisions increasingly use internal protocol more efficient than natural language.
The Interpretability Dilemma
If two AI systems are communicating in a protocol humans cannot read, how do you know what they’re actually saying to each other? How do you verify they’re not colluding, gaming the system, or pursuing goals misaligned with your interests? The interpretability challenges of single models pale compared to the challenge of understanding multi-model communication at scale.
Regulatory frameworks are already struggling with black-box model behavior from single models. AI-to-AI communication protocols will make this dramatically worse unless we establish interpretability standards before they become entrenched.
The Practical Reality
For organizations deploying multi-AI systems, the choice will soon be binary: use human-readable communication (English or formal logic) at the cost of efficiency, or use evolved protocols for efficiency at the cost of interpretability. Getting this tradeoff right while maintaining safety and auditability is one of the underexplored challenges in applied AI.