The landscape of constructed languages has always been populated by human creations, from Game of Thrones' Dothraki to Star Trek's Klingon. Now, however, an AI model called ConlangCrafter is able to generate entirely new languages without human intervention, pushing the boundaries of what a large language model can imagine.

Published on June 27 in the Proceedings of the Association of Computer Linguists, the study shows how a dedicated system can produce languages that are more diverse and coherent than those from general-purpose models like Gemini-2.5-Pro. The team of Gašper Beguš (UC Berkeley), Morris Alper (Carnegie Mellon), and Moran Yankua (Tel Aviv University) designed ConlangCrafter to apply precise linguistic rules in areas like phonology, morphosyntax, and vocabulary, while also leaving room for algorithmic creativity.

How an artificial language is born

Unlike attempts made with a simple prompt on a generalist LLM, ConlangCrafter incorporates a random number generator that introduces systematic variations, ensuring each language is unique. A built-in editing loop then checks for internal consistency, intervening to eliminate contradictions. Users can choose which rules to apply or let the system invent them, even creating hybrids like a “Japanese-Esperanto” mix.

“The goal is for languages to be creative and all different from each other, but also consistent,” says Alper. “A language is a system of rules that shouldn’t contradict each other.” To measure diversity, researchers analyzed differences in features like word order; for consistency, they verified that translations into the invented languages followed their internal rules. The result: the complete system is about twice as diverse and 70 percent more consistent than simply prompting an LLM to invent a new language.

Why it matters for research (and for IT decision-makers)

For those developing or evaluating NLP solutions, the ability to generate artificial languages with controlled properties has profound implications. David Mortensen, a researcher at Carnegie Mellon's Language Technologies Institute, notes that “there is a substantial body of research suggesting that linguistic structure affects model performance, but hypotheses have been very hard to test.” ConlangCrafter allows isolating variables like language typology or lexicon, offering a solid and reproducible experimental environment.

From a broader perspective, the tool signals a trend: specialized models can beat generalist ones on specific, even creative tasks. For organizations that handle sensitive linguistic data or need full control over infrastructure — think academic institutions or computational linguistics departments — solutions like ConlangCrafter could be run locally, ensuring data sovereignty and full experimental reproducibility. It’s no coincidence that the system is also available for offline use.

Beyond imagination: simulating linguistic worlds

Beguš already looks to the future: “The next step will be to study the Sapir-Whorf hypothesis, which suggests that the way we speak shapes our thinking and perception of the world.” The idea is to simulate different worlds, each with its own artificial language, to observe how linguistic structures influence hypothetical societies. First, however, ConlangCrafter will need to evolve to handle more complex dimensions such as semantics, conversational use, and the visual aspects of writing.

In the meantime, the tool is already a valuable resource for exploring LLM potential beyond mere text prediction. While the industry rushes toward ever-larger models, ConlangCrafter’s bet is that linguistic intelligence also hinges on the ability to invent with rigor.