AI Invented Entirely New Language, With Its Own Grammar, Vocabulary, and Sounds

Languages like Klingon and Elvish took their creators years, sometimes decades, to develop. Now researchers have built an AI system that can design the foundations of a new language in a single session, complete with its own sound system, grammar rules, and vocabulary. The system, called ConlangCrafter, doesn’t produce gibberish with foreign-sounding words. It generates languages meant to follow their own internal rules, the way real languages do, though the researchers found that keeping those rules perfectly consistent is still a challenge.

Creating a new language, known among enthusiasts as a “conlang” (short for constructed language), has long been a labor of love requiring deep expertise. Tolkien spent decades crafting Elvish for his fantasy world. The makers of Game of Thrones hired a professional linguist to build Dothraki. Esperanto, designed to bridge international communication, took years of careful philosophical and linguistic work. ConlangCrafter, developed by a team of researchers from Tel Aviv University, Carnegie Mellon University, and UC Berkeley and presented at the 2026 Association for Computational Linguistics meeting and published in the ACL Anthology, aims to make that entire process something a machine can handle, or at least give a serious head start to anyone who wants to try.

What sets this apart from a novelty is the system’s approach. Rather than asking an AI to dream up an entire language in one shot, which the researchers found produces dull, repetitive results, ConlangCrafter breaks the job into layers, building each piece of the language one step at a time and checking its own work along the way.

How ConlangCrafter Builds a Language From the Ground Up

ConlangCrafter’s core insight is that language has distinct layers, and those layers depend on each other in a specific order. Sounds come first: what consonants and vowels does the language use, and how are they allowed to combine? Then come grammar rules: how are sentences ordered, how do verbs change to reflect who’s doing what, and how is possession or time expressed? Only after those foundations are in place does the system fill in actual words.

At each stage, the AI generates a description of that layer, reviews its own work for contradictions, and revises before moving on. The researchers tested the system using several leading AI models, including DeepSeek-R1 and two versions of Gemini 2.5. This self-checking loop matters because the system is inventing rules as it goes, and small contradictions can snowball into bigger problems if left unchecked.

To keep things from getting repetitive, the researchers added a random number generator. Rather than letting the AI default to whatever language patterns it finds most familiar from its training data, which tends to produce languages that cluster around a limited set of familiar structures, the system randomly selects features like word order, how verbs are built, and whether the language uses tones. The AI then fills in the details based on those randomly chosen starting points, keeping creative control of the fine-grained design while letting randomness push it into less-explored territory.

How Researchers Measured Something That Has Never Been Measured Before

One of the trickiest parts of this research was figuring out how to grade the results. There’s no correct answer when you’re judging an invented language. No dictionary exists, no native speakers can weigh in, and no two grammars need to look alike. The team developed a new evaluation system to tackle this problem.

For measuring variety, the researchers drew on a real-world database called the World Atlas of Language Structures, which catalogs how natural languages around the world differ on key features: things like whether sentences typically put the verb before or after the subject, whether the language adds sounds before or after a root word, and whether it distinguishes grammatical gender. They selected 16 such features and used an AI judge to analyze generated languages against the same checklist. By comparing how differently the generated languages scored from one another, and from real languages in the database, they could put a number on variety.

To measure internal consistency, the team translated a fixed set of 10 test sentences into each generated language and had an AI judge check whether each translation followed the rules the language had established for itself. Sentences tested a range of grammatical situations, from basic statements to questions to commands.

Across 20 generated languages with 10 test sentences each, ConlangCrafter outperformed the simpler baseline approach on both measures. Variety scores for the baseline hovered around 0.25 to 0.35 depending on the AI model used, while ConlangCrafter consistently scored around 0.56 to 0.60. The real-world natural language sample from the database scored about 0.43 on this same 16-feature metric, meaning ConlangCrafter’s languages showed greater structural variety than the WALS sample by that specific measure. Consistency scores increased for all three models, though the improvement for Gemini 2.5 Flash was not statistically significant at the study’s sample size.

To check whether the automated scoring actually tracked expert judgments, the team brought in two PhD-level linguists, experts in language documentation and typology, who spent approximately 35 hours reviewing a subset of the generated languages and translations by hand. Agreement between the human experts and the automated scores was meaningful, with consistency evaluations showing a statistically significant positive relationship. The researchers note that this level of agreement is standard for creative evaluation tasks, where subjective judgment naturally varies even among trained experts.

ConlangCrafter Can Generate Languages No Human Would Ever Naturally Speak

Beyond the numbers, the system can do something genuinely unusual: generate languages built around constraints that no natural human language would ever have evolved to satisfy. In one example, a user specified that the language should have no consonant sounds at all, only vowels. The system produced one accordingly. In another, a user specified that the language was designed for an alien species and, instead of sounds, should use colors and gestures as its basic building blocks. The system invented what it called “chromemes,” color-based signals, and “kinemes,” gesture-based signals, and built a grammar around them.

These examples point toward real applications: game developers building fictional worlds that need internally coherent cultures and languages, or researchers testing theoretical ideas about what features a language actually needs to function. The system also lets users with partial ideas fill in what they know and have the AI handle the rest, much like a drafting program helps an architect who knows what a building should look like but needs help with the engineering.

Limitations the Researchers Flag

Researchers are candid that ConlangCrafter is a starting point, not a finished product. The languages it produces cover only the basics: sounds, grammar structure, and a vocabulary list, leaving out subtleties of meaning, how conversations flow, and how words are written down. The system’s vocabulary and grammar descriptions are also shorter than those in a real language reference, limited by how much text the underlying AI models can handle at once.

There are also concerns about bias. Large language models learn primarily from text in widely spoken languages like English, which may subtly push generated languages toward familiar patterns even when the system tries to diversify. The researchers also caution that generating fictional language content should not pull research attention or resources away from the urgent work of documenting real endangered languages spoken by living communities today.

Running the full pipeline is computationally expensive, with translation and self-checking loops accounting for most of that cost.

For anyone who has ever tried to invent a language and given up after three vocabulary lists, that may still sound like a bargain.

---