Researchers have shown that autonomous artificial intelligence (AI) systems can spontaneously develop a functional and shared language without human intervention. AI agents construct their own communication system that grows directly out of their specific sensors and environment.
This suggests advances for communication between autonomous systems such as self-driving cars, robots and industrial sensors. When AI agents are equipped with sensors for which humans have no direct equivalent - such as lidar, sonar, infrared or chemical sensors - existing human language is inadequate.
How the agents build their own vocabulary
The researchers provided the agents with basic mechanisms inspired by how human languages evolve: inventing words locally, adjusting conceptual prototypes, and learning from feedback when communication succeeds or fails. Rather than forcing AI systems to use human language that does not match their sensors, the researchers allow them to develop their own "artificial natural language." The agents build effective concepts that precisely match what their sensors perceive, even when that data is incomprehensible to humans.
A shared language can provide a layer of abstraction above raw sensor readings, and this research lets AI agents construct that abstraction autonomously.
No central computer imposes rules. Two agents communicate at a local level, and through these individual interactions a global convention emerges spontaneously across the entire population. This self-organisation lets the language optimise itself automatically, with agents creating exactly the number of categories needed for their tasks.
The agents successfully developed vocabulary to communicate about physico-chemical analyses of wine, NASA data on exoplanets, visual datasets, and images of birds or faces. In each case they synthesised data into new, shared concepts.
These systems offer notable transparency advantages over "black-box" AI models. Researchers can determine precisely what a developed word means because its prototype links directly to underlying sensor values. The linguistic systems are also robust: when environments change, sensors become noisy, or agents differ in structure, the language adapts automatically through continuous interaction.
This advance matters because it points toward AI systems that can collaborate and share information about experiences no human language can capture. Self-driving fleets could warn each other about road conditions measured through sensors humans lack. Industrial robots could coordinate tasks using concepts grounded in their unique perceptual capabilities. Rather than depending on human translation or rigid pre-programmed codes, machines can develop the linguistic tools they need for their specific worlds.