ConlangCrafter lets AI build languages with their own grammar

A research team has developed a system that can produce the foundations of a constructed language from almost nothing: a sound inventory, word-building rules, sentence order, vocabulary and translated examples. ConlangCrafter was presented at ACL 2026, a major conference in computational linguistics.

This is not a secret language that emerged spontaneously between machines. ConlangCrafter is a deliberately designed workflow that orchestrates pretrained language models. Its contribution lies in splitting a difficult language-design project into verifiable stages and trying to preserve consistency across them.

Sounds first, followed by grammar and vocabulary

The system begins with a “language sketch”, a reference document that grows during generation. It first defines the phonology: available sounds, possible combinations and pronunciation constraints. It then adds morphology and syntax before building an initial vocabulary compatible with those rules.

The order matters. Word forms need to follow the sound system, while example sentences must obey the chosen grammar. A single broad request to a model is more likely to produce a description that looks plausible but remains incomplete, repetitive or contradictory.

A second phase performs “constructive translation”. When a new sentence requires a missing word or rule, the system can propose the addition, store it in the language sketch and reuse it in later examples.

Randomness helps break the model’s habits

Without extra constraints, large models tend to return to structures that are common in their training material. ConlangCrafter therefore asks the model to prepare checklists of possible typological features. A random-number generator selects options before the phonology and grammar stages are written.

This can force less familiar combinations: an unusual order of subject, verb and object, a tonal system, different grammatical alignments or rare sound distinctions. The model then has to build around those choices instead of repeating its preferred pattern.

A critic finds problems and an editor revises them

After each stage, a refinement loop searches for ambiguity and contradiction. The same base model takes two successive roles. A critic identifies problems, then an editor rewrites the description. The cycle ends when no further issue is reported or when it reaches a fixed iteration limit.

The researchers tested the method with DeepSeek-R1 and the Flash and Pro versions of Gemini 2.5. They generated roughly twenty languages per configuration and used ten test sentences for each translation evaluation.

Greater variety without a guarantee of a perfect language

The study measures diversity across sixteen features derived from the World Atlas of Language Structures. On that specific measure, ConlangCrafter scores between 0.56 and 0.60, compared with 0.25 to 0.35 for a single-stage baseline. The calculated value for a reference set of 1,874 natural languages is 0.43.

This does not establish that the generated systems are generally richer than human languages. It shows that the sample explores a wider set of typological combinations across the sixteen selected dimensions.

Consistency remains the main limitation. Depending on the model, the automatically measured proportion of translations judged fully compliant with the language sketch reaches only 0.38 to 0.54. The pipeline improves on the baseline, but substantial descriptions can still contain rules that are difficult to apply or examples that conflict.

Two linguistics PhD students spent about 35 hours on a smaller manual assessment. Their work supports the usefulness of the automated measures while also showing how difficult it is to evaluate a newly designed language when no ground truth exists.

A creative tool and a test bed for linguistics

ConlangCrafter could accelerate fictional world-building, help a creator expand a grammar or produce unusual cases for testing a model’s metalinguistic reasoning. The authors also discuss possible applications to languages with limited digital resources, although that direction requires separate validation with the relevant communities and specialists.

The ConlangCrafter project website provides generated languages, sound systems and examples. The research paper describes the method, experiments and limitations.