Princeton researchers have trained AI on the speech surrounding individual children, finding that replaying recent experience was crucial to retaining what they heard.
The study, published November 5 in Science, used recordings from eight children in the university’s 1,000 Days Project. Each model encountered one child’s speech environment chronologically, without written text, a dictionary or predefined word boundaries.
Simply hearing successive days of speech proved insufficient in the initial model. Researchers then added replay: after each simulated day, the system revisited a limited sample of recent experience.
With replay, the model identified speech sounds and thousands of recurring words. The same settings also supported learning across seven other households, despite differences in their conversations.
The findings bring experimental detail to proposals that AI should learn more like children and research examining how memory shapes language learning.
These systems did not understand word meanings. They also lacked bodies, social motivations and the biological machinery of children.
Matching individual developmental timing required adjusting replay, learning speed and the onset of word learning. The result offers a way to test theories of language acquisition, rather than a complete artificial model of childhood.