Why 'go/went' persists: AI simulations reveal language evolution secrets

Why do languages keep irregular forms like 'go' → 'went' when regular patterns like 'walk' → 'walked' are easier to learn? A team from the University of Oregon and Google has tackled this question with a new computational approach combining agent-based modeling and large language models. In their preprint on arXiv (2606.12748), Aravinth Kulanthaivelu and Richard Sproat simulate how morphological alternations emerge and spread through populations. Their multi-agent system lets agents 'hear' novel forms (e.g., a different past tense for 'go') and adopt them probabilistically, allowing alternative forms to spread across the paradigm. Unlike prior work, their model supports naturalistic lexicons with hundreds or thousands of entries, realistic phonological rules, and populations of tens to hundreds of agents. They also test multiple network topologies, including scale-free social networks, which proved crucial for generating plausible alternation patterns.

The paper's key innovation is the AI Historical Linguist, an LLM-driven system that stages a debate between two simulated historical linguists. Given a set of morphologies—some real, some disguised, some evolved in simulations—the AI judges which are most realistic. This provides a quantitative evaluation method that goes beyond human intuition. Results indicate that random Bernoulli adoption of forms combined with scale-free social networks yields the most natural-looking outcomes. The authors also present three case studies of attested historical changes (e.g., the rise of 'went' in English), using the model to explore counterfactual scenarios. All code and data are released, enabling further research into the deep mechanisms behind language irregularities that have puzzled linguists for decades.