How did computer models challenge the idea that grammar requires an inborn language device?

The idea that children require an inborn "language acquisition device" to master grammar—a theory famously championed by Noam Chomsky—has been challenged by researchers using neural network computer models.1

While proponents of the language instinct argue that universal constraints on syntax and the "poverty of the stimulus" necessitate innate grammatical knowledge, opponents have used computer simulations to demonstrate an alternative. These models suggest that complex linguistic features can emerge from self-organization within a network without any predesigned rules or grammar. In these simulations, constraints on syntax and developmental timing arise as the network progressively adjusts its connection strengths through repeated feedback loops.1

A notable example of this approach is connectionist research, such as the 1986 experiment where a network was trained to form the past tenses of English verbs. By adjusting connection weights based on input-output differences, the network learned to generalize rules—correctly conjugating unfamiliar verbs—without being explicitly programmed with linguistic principles. While these studies provide a compelling counter-narrative to nativist theories, the debate remains unsettled, as the models do not fully account for all aspects of human linguistic development.21