AI agents develop their own 'artificial natural language' to communicate about complex data

edited by Swati Mestri, reviewed by Robert Egan

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources Evolutionary dynamics during the training phase of the CLEVR experiment with 10 agents. Mean shown as a line, shaded region indicates ±2 standard deviations over 10 runs. Credit: Transactions of the Association for Computational Linguistics (2026). DOI: 10.1162/tacl.a.799 Researchers from the VUB's Artificial Intelligence Lab and the University of Namur (UNamur) demonstrate in a new large-scale study how autonomous AI systems can spontaneously develop a functional, shared language without human intervention.

Unlike large language models such as ChatGPT, which memorize existing human language, these AI agents construct their own communication system, directly rooted in their specific sensors and environment.

This discovery represents a breakthrough for communication between autonomous systems, such as self-driving cars, robots and industrial sensors.

The research is published in the journal Transactions of the Association for Computational Linguistics.

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 equivalent—such as lidar, sonar, infrared or chemical analysis—existing human language proves insufficient.

Jérôme Botoko Ekila, a doctoral student at the VUB's AI Lab, explains, "Instead of forcing AI systems to communicate in a human language that does not correspond to their sensors, we allow them to develop their own 'artificial natural language.' They construct effective concepts that correspond exactly to what they perceive via their sensors, even if the data is incomprehensible to humans."

Because sensors on different machines or vehicles can never be calibrated identically in practice, the researchers believe that this language offers the ideal solution.

Paul Van Eecke, a professor of artificial intelligence at the VUB, says, "In industrial or physical applications, it is simply unrealistic to expect all the sensors on different robots or vehicles to be calibrated in exactly the same way.

"Humans solve this problem through language: if I say I am standing near the 'red car,' we understand each other without our eyes registering exactly the same color values. Language provides a layer of abstraction above the sensor readings, and this is precisely what we are now allowing these AI agents to construct autonomously."

Instead of preprogrammed words or fixed codes, the researchers provided the agents with the basic mechanisms underpinning the evolution of human language: the local invention of words, the adaptation of conceptual prototypes and the incorporation of feedback following successful or failed communication.

Katrien Beuls, a professor at UNamur, explains, "What's fascinating is that there's no central computer imposing the rules. Each time, we let two agents communicate with one another at a local level. It is solely through these individual interactions and updates that a global convention emerges spontaneously across the entire population.

"This principle of self-organization allows the language to optimize itself automatically: the agents create exactly the number of categories needed to carry out their tasks efficiently, without superfluous details."

The universal applicability of the method has been demonstrated across 37 varied and complex data sets. The agents successfully developed their own vocabulary to communicate, notably regarding physicochemical analyses of wine, NASA data on exoplanets, visual data sets and images of birds or faces. In all these situations, the agents successfully synthesized the data into new, shared concepts.

Transparency is a notable advantage of the systems developed. Unlike many "black-box" AI models, these systems allow researchers to precisely determine the meaning of a specifically developed word, as the prototype can be directly linked to the values of the underlying sensors.

Furthermore, these evolving linguistic systems are robust. When the environment changes, when sensors produce noise or when individual agents have a different structure, the language automatically adapts to the new circumstances through continuous interaction.

Jérôme Botoko Ekila et al., Emergent Communication in Continuous Worlds: Self-Organisation of Conceptually Grounded Vocabularies at Scale, Transactions of the Association for Computational Linguistics (2026). DOI: 10.1162/tacl.a.799 Key concepts Large language models AI alignment Human-centered AI interfaces Agentic consumer AI Provided by Free University of Brussels Who's behind this story? Swati Mestri Swati Mestri holds a bachelor's degree in Electronics Engineering and has worked as a content editor since 2019. She has experience editing research documents across technology, health care, and materials science, and has a particular interest in technology and space. Full profile →

Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →