New research from the University of the Witwatersrand in South Africa reveals how iterated learning drives the evolution of structured language in both children and artificial intelligence. Published in the Proceedings of the National Academy of Sciences, the study, led by Dr. Devon Jarvis, demonstrates that language becomes more systematic over generations because it adapts to the constraints of how learners process information. The team developed a computer model mimicking a child's cognitive development to observe how language structures emerge. They fed the model data with properties similar to human language and tracked how successive generations of the model learned and transmitted information. The results showed that the computational system identified structural patterns in the data much like human children. Crucially, the dataset itself became increasingly structured over time, proving that this evolution occurs because it simplifies the learning process for subsequent generations. Jarvis explains that children acquire knowledge hierarchically, starting with basic concepts before moving to complex ones. For example, a child may learn that birds fly and then over-generalize this rule to penguins, initially failing to understand that penguins cannot fly. These errors are not random; they stem from over-extrapolating existing knowledge. Through these mistakes and subsequent corrections, children refine their understanding. In the context of language transmission, parents pass information to children, and these transmission errors act as a filter. Easy-to-learn portions of language are retained and reused, while unstructured or difficult elements are discarded. Consequently, language naturally evolves toward structures that align with human learning biases. To test these hypotheses, the researchers utilized deep linear neural networks, which simulate brain processing. They found that the phenomenon of iterated learning only succeeded when the network possessed sufficient depth and complexity. Shallow networks with fewer layers failed to capture the regularities that make language learnable. This suggests that the architecture of a learning system and the richness of its environment are critical factors in absorbing and transmitting language structure. The findings have significant implications for the field of artificial intelligence. As generative AI models rely heavily on scale to develop emergent capabilities, this study highlights that the specific nature of how learners acquire information is fundamental to language evolution. Jarvis notes that while the components of this research, such as deep linear networks and iterated learning, have existed in separate literatures for years, combining them reveals a unified principle. The fact that even a simplified version of the technology behind modern AI tools exhibits these cognitive behaviors suggests that the intersection of biology and computation holds the key to understanding the fundamental principles of human cognition.