Can autonomous AI agents spontaneously develop a functional, shared language without human intervention? The answer is YES. Researchers from the Artificial Intelligence Lab at the VUB and the University of Namur (UNamur) have demonstrated this in a new large-scale study.
Unlike Large Language Models such as ChatGPT, which learn existing human language through pattern recognition, AI agents build their own communication system that is directly anchored in their specific sensors and environment. This finding represents a significant step forward for communication between autonomous systems, such as self-driving cars, robots and industrial sensors.
A layer of abstraction for non-human senses
Current AI language models focus on understanding and generating human languages such as Dutch or English. These human languages are historically rooted in our own sensory perceptions. When physical or virtual AI agents are equipped with sensors for which humans have no direct sensory counterpart – such as lidar, sonar, infrared or chemical analysis – existing human language falls short.
AI agents’ own language
Jérôme Botoko Ekila, a researcher at the VUB’s AI Lab, explains: “Instead of forcing AI systems to communicate in a human language that does not suit their sensors, we allow them to evolve their own ‘artificial natural language’. They build effective concepts that are precisely tailored to what they perceive through their sensors, even when the data is incomprehensible to humans.”
As sensors on different machines or vehicles can never be calibrated identically in practice, the researchers believe that language offers the ideal solution.
Abstraction layer
Paul Van Eecke, professor of artificial intelligence at the VUB: “In industrial or physical applications, it is simply unrealistic to expect that all the sensors on different robots or vehicles are calibrated perfectly identically. People solve that problem with language: if I say I’m standing by the ‘red car’, we understand each other without our eyes registering exactly the same colour values. Language forms a layer of abstraction above sensor values, and that is precisely what we are now letting these AI agents build autonomously.”
Self-organisation and evolution
Instead of pre-programmed words or fixed codes, the researchers simply provided the AI agents with the basic mechanisms underlying human language evolution: the local invention of words, the adaptation of conceptual prototypes, and the incorporation of feedback following successful communication or errors.
From individual interaction to global convention
Katrien Beuls, a professor at UNamur, explains: “The fascinating thing is that there is no central computer imposing the rules. We allow two AI agents to communicate with each other at a local level each time. Purely through these individual interactions and updates, a global convention emerges spontaneously across the entire population. This principle of self-organisation ensures that the language optimises itself automatically: the agents create exactly enough categories to carry out their tasks efficiently, without superfluous details.”
Tested on 37 datasets
The universal applicability of the method was demonstrated across 37 diverse and complex datasets. For example, the agents successfully developed their own vocabulary to communicate about, amongst other things, physicochemical analyses of wine, NASA data on exoplanets, and visual datasets such as images of birds or faces. In all these situations, the agents succeeded in efficiently summarising data into new, shared concepts.
Transparent and adaptive
A notable advantage of the systems developed is their transparency. Unlike many ‘black box’ AI models, researchers can determine exactly what a specific word developed means, as the prototype can be directly traced back to the underlying sensor values.
Furthermore, these evolutionary language systems are extremely robust. When the environment changes, when sensors produce noise, or when individual agents have different structures, the language automatically adapts to the new circumstances through continuous interaction.
Funding and collaboration
This research was carried out as part of a collaboration between the Artificial Intelligence Laboratory at the Vrije Universiteit Brussel (VUB) and the Faculty of Computer Science at the Université de Namur (UNamur), and was funded by the Research Foundation – Flanders (FWO) and the Fonds de la Recherche Scientifique (FNRS) through a WEAVE project.
Photo: Image by GenerativeStockAI via Pixabay; Source: VUB
See also: Artificial Intelligence: Dutch industry is still missing out on many opportunities
