A large-scale study by the Artificial Intelligence Lab at the VUB and the University of Namur (UNamur) demonstrates that autonomous AI systems can spontaneously develop a functional, shared language. Unlike large language models trained on human text, these AI agents construct their own communication systems, directly grounded in data from their specific sensors and environment. This breakthrough is significant for communication between autonomous systems like self-driving cars, robots, and industrial sensors.
Current AI language models focus on human languages, rooted in human sensory perceptions. However, when AI agents use sensors without direct human equivalents (e.g., lidar, sonar, infrared, chemical analysis), human language proves insufficient. Jérôme Botoko Ekila, a doctoral student at the VUB AI Lab, explains: "Instead of forcing AI systems to communicate in a human language that doesn't match their sensors, we allow them to develop their own 'artificial natural language'." These languages create effective concepts corresponding to their perceptions, even if incomprehensible to humans.
Language serves as an ideal solution when sensors from different machines are not identically calibrated. Paul Van Eecke, professor of artificial intelligence at the VUB, notes: "In industrial or physical applications, it is unrealistic to expect all sensors across different robots or vehicles to be calibrated identically. Humans solve this with language: if I say I'm near the 'red car,' we understand each other without our eyes registering the exact same color values. Language is an abstraction layer above sensor values, and that's precisely what we're now letting these AI agents build autonomously."
Instead of pre-programmed words or fixed codes, researchers provided agents with basic mechanisms underlying human language evolution: local invention of words, adaptation of conceptual prototypes, and feedback on communication success or failure. Katrien Beuls, professor at UNamur, explains: "What's fascinating is that there's no central computer imposing rules. We let two agents interact locally. Only through these individual interactions and updates does a common convention emerge spontaneously within the entire population." This self-organization principle allows the language to optimize automatically, creating exactly the necessary categories without superfluous details.
The method's universal applicability was demonstrated across 37 diverse datasets. Agents successfully developed vocabularies for communicating about physico-chemical wine analyses, NASA exoplanet data, visual datasets, and images of birds or faces, effectively grouping data into new common concepts.
These evolving linguistic systems offer notable transparency. Unlike many 'black box' AI models, researchers can precisely determine the meaning of an agent-created word because the prototype directly links to underlying sensor values. Furthermore, these systems are robust; the language automatically adapts to environmental changes, sensor noise, or differing agent configurations through continuous interaction.
This research, published in the Transactions of the Association for Computational Linguistics, is a collaboration between the Vrije Universiteit Brussel (VUB) and the Faculty of Computer Science at the University of Namur (UNamur), funded by the FWO and the FNRS.
