Artificial intelligence is stepping beyond translation and into invention, using large language models to create brand-new humanlike (and even radically nonhuman) languages. A recent paper in the Proceedings of the Association for Computational Linguistics describes ConlangCrafter, a tool designed to generate constructed languages with explicit grammar rules, vocabularies, and sound systems.
Instead of asking a general model to “make a language” in one shot, ConlangCrafter uses a multi-hop pipeline that breaks the task into subproblems. The system first proposes a phonological inventory (or deliberately removes it), then builds word-formation constraints, and finally specifies syntax and sentence-building behavior.
A key ingredient is iterative self-checking. After generating a language, ConlangCrafter translates sentences from natural languages into the constructed one, then reviews the output against its own “language sketch”—a living specification of rules. When inconsistencies appear, the pipeline revises both the sketch and the translations, improving internal consistency rather than relying on stylistic plausibility alone.
The researchers report that their generated languages are both more diverse and more internally coherent than languages produced by prompting a general-purpose LLM to invent one from scratch. In tests, ConlangCrafter is evaluated for how consistently translations obey the constructed grammar and for how much the languages vary in features such as unique sounds and structural patterns.
ConlangCrafter is also parameter-driven, enabling controlled experiments in linguistic extremes. For example, the tool can create a language with no consonants at all, or generate an “alien” communication system for a cephalopod species that relies on colors and gestures rather than speech.
While this may sound like pure creative writing, the authors emphasize research value. Constructed languages could help linguists explore hypotheses about language evolution, test assumptions about grammar under constrained conditions, and study how AI agents coordinate communication when no shared natural language exists.
A major challenge the team tackled was evaluation: defining objective metrics for something inherently creative. Their framework quantifies rule-following behavior and measures diversity across linguistic feature sets, allowing the pipeline to be tested in a way that resembles scientific experimentation.
ConlangCrafter was released recently and has already drawn attention in mainstream technology coverage, highlighting the growing momentum behind AI systems that can generate structured symbolic artifacts—not just text, but rule-governed communicative systems.
Subject of Research: Experimental study
Article Title: ConlangCrafter: Constructing Languages with a Multi-Hop LLM Pipeline
News Publication Date: 2-Jul-2026
Web References: https://aclanthology.org/2026.acl-long.422.pdf ; https://conlangcrafter.github.io/ ; https://www.science.org/content/article/ai-can-invent-entirely-new-languages-it-creative ; https://spectrum.ieee.org/conlangs-ai-model-contructed-languages
References: Proceedings of the Association for Computational Linguistics (ACL) paper on ConlangCrafter
Image Credits: Not specified
Keywords: artificial intelligence, large language models, conlang generation, phonology, syntax, linguistic evaluation, machine translation, structured language creation
