ConlangCrafter: AI Learns to Create Constructed Languages
UC Berkeley
Google/DeepMind
Researchers from the University of California, Berkeley have developed the ConlangCrafter model, capable of generating new constructed languages (conlangs) with high diversity and consistency. The system outperforms standard LLMs such as Gemini-2.5-Pro by nearly double on these metrics, and can be applied to study non-human languages and test the influence of linguistic structure on NLP model performance.
In a paper published on June 27 in the Proceedings of the Association for Computational Linguistics, a model called ConlangCrafter is described, developed under the leadership of Gasper Begus (University of California, Berkeley) together with Morris Alper (Carnegie Mellon University) and Mor Yanuka (Tel Aviv University). ConlangCrafter can create artificial languages, following linguistic rules specified by the user (phonology, morphosyntax, lexicon) or generating them on its own. A random number generator introduces variability, and a built-in editing loop fixes inconsistencies. The system can generate unusual languages, for example, for cephalopods using colors and gestures. When compared to general-purpose large language models such as Gemini-2.5-Pro, ConlangCrafter showed roughly twice the diversity and nearly 70% greater consistency. David Mortensen from Carnegie Mellon University noted that the model could help NLP researchers study the impact of linguistic structure on model performance. ConlangCrafter is available for free online; the developers acknowledge limitations in semantics, contextual language use, and visual aspects of writing systems. Begus plans to use the model to test the Sapir-Whorf hypothesis about the influence of language on thought.
Source: IEEE Spectrum AI —
original