AI tool builds entirely new languages, then checks its own grammar

An artificial language with no consonants sounds like a puzzle. A language spoken through colors and tentacle gestures sounds stranger still. ConlangCrafter can build both, while trying to keep every rule, word, and translation consistent.

The system uses large language models to create constructed languages, or conlangs, one layer at a time. Morris Alper, Moran Yanuka, Raja Giryes, and Gašper Beguš describe the method in the Proceedings of the Association for Computational Linguistics.

“The idea of ConlangCrafter is: ‘How can we make new languages that have different linguistic features than what we normally see in natural languages?’” said Alper, the paper’s first author and an incoming University of Miami computer science professor.

The team has created more than 60 languages and released the code online. Users can also set unusual constraints before generation begins.

A general-purpose model may produce something language-like after one prompt. The result, however, can contain mismatched sounds, unclear grammar, or vocabulary that changes without warning.

“Imagine if you just say, ‘Make me a language.’ It will give you something that doesn’t make sense,” Alper said. “What we did is build this pipeline where you say, ‘Okay, what are the sounds? And then let’s check them. And then, what are the rules for building words? What about syntax?’ We split the problem apart and have the LLMs solve each sub-problem and combine them together.”

ConlangCrafter divides the job into phonology, grammar, and vocabulary. Phonology covers a language’s sound system. Grammar includes word formation and sentence structure. The lexicon contains the language’s words.

The tool builds these layers in sequence because later choices depend on earlier ones. Its “language sketch” records rules, examples, and additions as the system works.

Afterward, the tool performs what the authors call constructive translation. It translates sentences, adds missing words or rules when needed, and records them for later use.

Large language models know many linguistic patterns, but they may return similar designs across repeated attempts. ConlangCrafter addresses that problem with structured randomness.

During sound-system and grammar development, the model creates checklists of linguistic features. A random number generator selects among possible options, pushing each language toward different sounds, word orders, and grammatical structures.

That process produced an all-vowel language with no consonant phonemes. Another language, designed for an alien cephalopod species, replaced speech sounds with colors and gestures.

In that example, color-based “chromemes” can mark qualities such as warm, cool, static, or pulsating. Gesture-based “kinemes” can involve curling tentacles. Movement patterns through water can work like tones.

The method also uses self-refinement. One model role identifies contradictions or unclear details. Another revises the language. The cycle continues until no further problems appear or the system reaches a preset limit.

Evaluating a fictional language creates an unusual problem. There is no reference translation or established grammar against which every result can be checked.

“The hardest part of the work was defining an objective measure to give us numbers that say, ‘How well is the model performing at this task?’” Alper said. “That’s really hard to do for creative things.”

The team measured typological diversity and internal consistency. Typology compares structural features across languages, including basic word order, sound inventories, tone, and grammatical patterns.

For diversity tests, the researchers used 16 features from the World Atlas of Language Structures. They compared generated languages with 1,874 natural languages that had enough recorded information.

Natural languages received a diversity score of 0.43. Baseline, single-prompt methods scored between 0.25 and 0.35. ConlangCrafter scored between 0.56 and 0.60, depending on the underlying model.

Those results suggest the system generated combinations that differed more widely than baseline outputs. Its languages also exceeded the measured diversity of the natural-language sample.

The researchers also tested whether translations followed each invented language’s rules. They generated about 20 languages per model and translated 10 test sentences for each one.

ConlangCrafter improved consistency across all three tested models. With Gemini 2.5 Pro, the score rose from 0.32 for the baseline to 0.54. DeepSeek-R1 rose from 0.21 to 0.39. Gemini 2.5 Flash increased from 0.30 to 0.38, although that difference was not statistically significant at the study’s sample size.

Ablation tests showed which parts mattered. Multi-step generation and random feature selection increased diversity. Those changes could make languages harder to keep coherent. Self-refinement recovered much of the lost consistency.

The team compared structured randomness with raising a model’s sampling temperature. Higher temperature increased variety, but consistency fell sharply. The checklist method reached 0.60 diversity with 0.43 consistency. High-temperature sampling approached that diversity but dropped consistency to 0.34.

Two linguistics doctoral students reviewed generated languages and translations during about 35 hours of work. Automatic feature extraction reached 91 percent accuracy when underspecified cases were excluded. Automated consistency scores showed moderate agreement with expert judgments.

Fully consistent translation remains difficult. Open-ended user requests are also hard to score because they lack one correct outcome. The current system omits semantics, pragmatics, and writing systems, which could appear in later versions.

ConlangCrafter could reduce repetitive work for people designing languages for films, television, books, and video games. It may help generate vocabulary, test grammar choices, or expand partial sketches while leaving users in control.

The method could also support research on poorly documented languages when detailed grammars exist but large text collections do not. Its rule-focused translation process may help test language technology under those conditions.

Other possible uses include studying how languages change and how artificial agents might develop communication systems over time. Because each target language begins as something new, the task also tests whether AI can reason through unfamiliar rules instead of relying only on memorized examples.

Research findings are available online in the journal Proceedings of the Association for Computational Linguistics.

The original story "AI tool builds entirely new languages, then checks its own grammar" is published in The Brighter Side of News.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.