American researchers develop a model that generates complete languages with their own grammar and phonology

Researchers in the United States have created a system that produces complete languages, including a sound system, grammar and vocabulary, rather than merely imitating existing patterns. Led by Professor Gasper Begus, the tool has already generated more than 60 distinct languages, among them a consonant-less language and a language built on colors and motions designed for fine-grained head-to-foot coordination.

The model operates in three stages. First it defines a phonology; on that foundation it constructs a grammar; finally it expands a lexicon. At each stage the system checks for internal contradictions and corrects them, allowing the rules to stabilize gradually without human intervention. Begus explained that users can "give a language any color you want" and demonstrated a blend of Japanese and Esperanto.

The researchers benchmarked the output against languages that Gemini 2.5 Pro produces from a single prompt starting from scratch. By their metrics, the dedicated model’s languages are twice as diverse on a structural-diversity measure, and internal consistency—assessed through grammatical and lexical contradiction tests—is higher by roughly 70 percentage points.

Beyond entertainment applications such as film, literature and gaming, the tool enables exploration of how linguistic structures might evolve under different conditions, for example in the absence of consonants or when the communication channel is visual. This opens a new avenue for studying the evolution of ancient languages without relying solely on fragmentary documentation.

The code and example datasets are available on GitHub, allowing researchers and creators to run the system themselves and extend it.