ConlangCrafter Turns AI to Imagining Languages
IEEE Spectrum Michelle Hampson
Researchers built an AI called ConlangCrafter that invents entire new languages from scratch, sounds, grammar and all. It's way more consistent than just asking ChatGPT to make up a language, and it might help us study how AI actually thinks about linguistic structure.
Based on reporting by IEEE Spectrum, Michelle Hampson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Humans have been inventing languages for fun and profit for a long time now, whether it's Klingon for Star Trek fans or Dothraki for Game of Thrones nerds who wanted more than subtitles. Now there's a new player in the conlang game, and it's not human. ConlangCrafter, built by a team including UC Berkeley linguist Gašper Beguš, generates full constructed languages on its own, complete with sound systems, grammar rules, and vocabulary, and it does it with a level of internal consistency that beats simply prompting a general chatbot to wing it.
The details matter here. ConlangCrafter doesn't just spit out random words. It builds phonology, morphosyntax, and lexicon in tandem, using a random number generator to keep every output distinct, then runs an editing loop that hunts down contradictions and patches them. Ask it for a mashup of Japanese and Esperanto, and it'll actually try to build something coherent rather than a word salad with borrowed grammar bolted on. The team even had it dream up a language for a hypothetical color-and-gesture-based cephalopod species, which isn't real octopus communication but is a genuinely interesting thought experiment about what language looks like outside a human mouth.
The numbers back up the ambition. When Beguš, Morris Alper of Carnegie Mellon, and Moran Yanuka of Tel Aviv University benchmarked their system against Gemini-2.5-Pro simply told to invent a language, ConlangCrafter came out roughly twice as diverse across features like word order and nearly 70 percent more internally consistent. That gap is the whole point: a made-up language is only interesting if it's actually a system, not just vibes and neologisms.
Outside researchers see a practical use beyond linguistic curiosity. David Mortensen at Carnegie Mellon, who wasn't part of the project, points out that natural language processing has long suspected that a language's structure shapes how well AI models learn and perform, but testing that cleanly has been brutal because real languages come tangled up with culture, data availability, and history. A tool that generates clean, rule-bound languages on demand could let researchers isolate variables in a way that's been nearly impossible until now.
Beguš, for his part, is already looking past the current version, which admittedly still can't handle semantics, conversational nuance, or writing systems. He wants to use future iterations to poke at the Sapir-Whorf hypothesis, the old idea that language shapes thought, by simulating entire societies built around different invented tongues. It's a strange, ambitious pitch: use AI's imagination to test theories about human imagination that we've never been able to properly isolate before.
My take — AI-written commentary, not fact-checked reporting
This is one of those rare AI papers that isn't chasing a benchmark score for its own sake, and I like that. Everyone's obsessed with whether models can pass the bar exam or write Shakespeare pastiche, but here's a tool that's actually useful for studying how structure and cognition interact, something linguistics has argued about for a century without good tools to test it. My only gripe is the framing that this is somehow more 'creative' than what humans do with conlangs, when really it's just faster and more rule-consistent, which is exactly the kind of grunt-work automation AI is actually good at.
Read more about this at: IEEE Spectrum
