AI tries, but humans still ace language learning

New research reveals that artificial intelligence develops more efficient communication by learning from human language structures rather than creating its own from scratch.

Artificial intelligence may be evolving quickly, but when it comes to developing a language, it still has a lot to learn from humans.

A study from the University of Gothenburg in Sweden shows that AI systems communicate more efficiently when they mimic human language patterns rather than inventing their own. Despite the growing sophistication of reinforcement learning, human language – which is shaped by culture, cognition, and centuries of evolution – continues to outperform artificial systems in terms of clarity, adaptability, and efficiency.

The research, based on two scientific papers included in a doctoral thesis, tested how well AI agents could create functional communication systems on their own. The agents were trained to play "signaling games" in which they had to describe either colors or numbers to one another. This is how they simulated a simplified version of language learning.

For instance, in one experiment, the sender agent saw a specific color chip from the Munsell color chart, a standard used in the World Color Survey. This agent had to describe it to a listener agent, who then had to guess which color it was meant to describe. The goal was simple: communicate successfully with as few misunderstandings as possible.

Using reinforcement learning algorithms, the agents were able to invent their own artificial "languages" – systems of signals that helped them complete the tasks. These systems showed surprisingly high levels of communication efficiency, which came close to the performance of real-world human languages.

However, when these artificial systems were compared with data from actual human languages, such as those documented in the World Color Survey and studies of numeral systems, the AI still fell short.

The key finding? The best-performing AI agents weren’t the ones that built a communication system from scratch, but those that aligned with the structure and statistical tendencies found in human language. The agents essentially learned better when they imitated the way humans naturally use words and categories.

“Human communication strategies provide a kind of efficiency benchmark,” said the lead researcher behind the study.

“Even without the full complexity of human perception and cognition, AI agents that learn from human patterns come closer to efficient communication than those left to build languages entirely on their own.”

How did AI do when it was noisy?

AI agents were trained in noisy or uncertain conditions, meaning the scientists were simulating some distractions or interference. This resulted in AI agents developing simpler communication systems.

This echoes a trend observed in human languages, where high-noise or stimulating environments often result in more compact or reduced forms of expression. For example, speakers might rely more heavily on context clues or use fewer words in environments where time or attention is limited.

The researchers also pointed out that many challenges remain. The AI systems studied lacked generalization skills – the ability to apply their language to new tasks or environments.

They were also unable to use compositional language, an aspect of human communication where meanings can be built from smaller parts, for example, “red” + “apple” = “red apple”. These limitations show that while AI can get close to mimicking human-like communication, it still doesn’t come with the flexibility or creativity that characterizes human speech.

For now, though, the conclusion is clear: the best way for AI to learn language is by taking cues from the humans who’ve already mastered it. Natural language remains not only more efficient but also more robust, adaptable, and generalizable than anything AI has been able to invent on its own.

So while machines are getting better at talking, they still talk best when they learn from us, humans.