AI learns languages similarly to humans, study shows

An AI system that learns language autonomously develops a language structured in the same way as human language. And just as we humans learn from previous generations, AI models get better when they take advantage of the knowledge of older relatives.

This is shown by a study from Chalmers University of Technology and the University of Gothenburg, Sweden, which explores the mechanisms behind human language and provides important knowledge for the development of the AI of the future.

The study is published in the Journal of Language Evolution.

AI-based language models like ChatGPT are getting better and better at mimicking human language and are increasingly being used as a tool for producing text. But the same types of models can also teach us to better understand how human languages have evolved.

In the study, the researchers explored an AI system for evolving languages by using two different methods in a previously untested combination. On the one hand, reinforcement learning was used—where the right actions are rewarded and thus reinforced—and on the other hand, the researchers let the AI models learn from each other over generations.

“We discovered that the AI models reached something that is structured in the same way as human language, and that language learning worked in a similar way to human language. Thus, the study provides important insights about how AI models work, but also a greater understanding of how human languages evolve,” says Emil Carlsson, who at the time of the study was a doctoral student at Chalmers and the University of Gothenburg.

All languages strive to be efficient

According to an influential theory in cognitive science, all human language is shaped by the need to be able to communicate effectively. At the same time, a balance is needed: as a tool, the language must be informative but also simple enough for us to learn. And the more information that needs to be communicated, the more nuanced the language needs to be.

A classic example is that languages in colder climates often have more words for snow and ice than languages in warmer climates.

To test the theory and investigate how language becomes efficient, the researchers created AI agents—different AI models—that played a communication game with each other.

The AI agents were shown a color and a list of symbols that initially had no meaning. As the agents interact, these symbols come to be associated with certain colors and are used to communicate the color to the other agent.

“The reason we used colors is that there is so much data on how the color spectrum is named in different languages, including data from isolated languages that have never been exposed to other languages. The categorization of colors varies between languages, both in terms of the number of words and which part of the color spectrum the words describe,” says Carlsson.

Rewards and generational exchange yielded results for language development

The experiments involved one AI agent communicating a color via one of the symbols in the list, and the receiving AI agent would guess which color the symbol corresponded to. Both agents received a common reward when they made progress in their communication. The closer they got to a common designation of the exact color shade that the receiving agent got, the more points were awarded.

In the next step, new “generations” of AI agents were created, while the old AI agents were phased out. The new AI agents got to see the dialogue and the language that the previous generation had managed to develop. After that, the new AI agents got to play the same communication game with each other.

“The idea was to let the AI agents first learn a language from previous generations and then further develop it by communicating with each other. Just like two small children who learn by listening to mum and dad talk and then continue to broaden and develop their own languages,” says Carlsson.

Provides knowledge about how language develops

The result was a system for naming colors that were similar to human color languages, despite the fact that the AI agents had never come into contact with them.

“The interesting thing was that it was precisely the combination of the problem-solving in the game, together with the fact that the AI agents learned from previous generations, that led to effective language that resembles human language. When the AI agents only communicated with each other to solve the game, the languages became too complex,” he says.

“We also tried to let the AI agents only learn from previous generations, without having to deal with the problem-solving aspect of the game, and then the languages became far too simple.”

According to Carlsson, the results indicate that our ability to communicate and learn from each other is crucial for how languages develop over time.

“When we only learn something from another person, without perhaps understanding the benefits of it, our tendency to develop the knowledge decreases. But when we actually have to use what we have learned to solve problems and move forward, that’s when structured and effective languages can be created,” he says.

He hopes that the results will contribute new insights and ideas in language research, as well as research in AI and computer science.

“This is knowledge that can help us better understand the mechanisms behind human languages, but also understand how large AI-based language models work. This can pave the way for being able to guide the development of AI in fruitful directions,” says Carlsson.