A new AI system can generate entire languages from scratch, complete with phonology, morphology, syntax, and lexicon. The question is whether that counts as creative.
ConlangCrafter, developed by a team led by linguist Gasper Begus at the University of California, Berkeley, uses a multi-hop LLM pipeline to produce constructed languages (conlangs) with consistent internal grammar rules. The work was presented as an oral paper at the ACL 2026 conference and has sparked a debate that goes to the heart of what it means to be creative.
How it works
ConlangCrafter is not a single trained model but a modular pipeline. It decomposes language design into sequential stages: phonology (sound system), morphology (word structure), syntax (sentence structure), lexicon generation, and constructive translation. At each stage, random number injection provides diversity, and self-refinement loops enforce internal consistency.
The system costs roughly $4 per language to run on DeepSeek-R1. The generated languages include features rarely seen in natural languages, click consonants, polysynthetic morphology, OVS (object-verb-subject) and VSO word orders, evidentiality marking systems, and ergative-absolutive alignment. One language was designed to use color changes and tentacle gestures, tailored for cephalopod-like communicators.
Begus, who also leads the linguistics team at Project CETI (which aims to decode sperm whale communication), sees ConlangCrafter as a tool for exploring the space of possible human languages and testing linguistic theory.
The creativity question
The debate turns on a fundamental question: is creating a language an act of creativity, or is the AI simply shuffling combinatorial options within a predefined space?
Joseph Windsor of the University of Calgary is the primary skeptic. “I don’t think we’d call the dice creative,” he told Science AAAS. Windsor argues that the generated languages lack consistency for long-term use and, crucially, lack the aesthetic feel that human conlangers bring to their work.
Christine Schreyer of the University of British Columbia, who has studied the living conlang community around the Na’vi language from James Cameron’s Avatar, points out that real language evolution requires community use. “Na’vi speakers coined their own word for ‘LOL'” she noted, something no AI system can replicate.
Balthasar Bickel of the University of Zurich adds that ConlangCrafter’s languages have no mechanism for diachronic change. They are born fully-formed and static, unlike natural languages and even many successful conlangs that evolve organically through use.
But other researchers see a different picture. If human creativity is itself combinatorial, selecting and recombining existing elements within constraint spaces, then the boundary between human and machine creativity may be blurrier than it appears.
“Why can’t an AI be called creative?” one researcher asked. Begus himself argues that creativity does not require consciousness. “The output is novel, structurally coherent, and follows rules,” he said. “If that’s the bar, ConlangCrafter clears it.”
Broader context
The ConlangCrafter debate arrives alongside other work on machine creativity. A 2026 study by Bellemare-Pepin and colleagues found that AI systems beat the average human on standard creativity tests, though the top 10% of humans still outperformed.
For linguists, the deeper question may be less about creativity and more about whether these generated languages can serve as productive tools for linguistic research. A system that can rapidly generate languages with specific grammatical features could help test universal grammar hypotheses, study the boundaries of human learnability, and explore the theoretical space of possible human languages.
Whether that counts as creative may ultimately depend on the definition used. And that definition, as the debate around ConlangCrafter shows, remains very much in flux.
Sources
Begus G, Alper M, Yanuka M, Giryes R. “ConlangCrafter: Multi-hop LLM pipeline for constructed language generation.” Proceedings of ACL 2026 (Oral). arXiv: 2508.06094
O’Donnell J. “An AI can invent entirely new languages. But is it creative?” Science (July 13, 2026). https://www.science.org/content/article/ai-can-invent-entirely-new-languages-it-creative