IASC system from Sakana AI creates languages from scratch without human participation

Imagine: previously, to create a language from scratch, you had to be Tolkien or a team of linguists from “Star Trek”. Years of work, thousands of hours on phonetics, grammar, writing. And now AI does this. Simply takes and assembles a language piece by piece, like a constructor.

Researchers from Sakana AI and University of Notre Dame built an IASC system that constructs languages step by step. First the model invents sounds and their combinations. Then grammar, declensions, word order. Further – writing. And for dessert outputs a mini-textbook with examples and translations. All by itself, without humans.

But there’s a nuance. The model handles even abstract things like “inclusive we” – you know, when the language has a difference between “we with you” and “we without you”. But as soon as it encounters something rare, for example non-standard word order like in Basque, and that’s it – AI stalls.

The work’s authors say this will help compose languages for games and films. But the most interesting thing is different: now scientists can check what AI actually understands about language. Where it really grasps concepts, and where it just reproduces statistics from training data. It turns out researchers created a tool to understand how much the previous tool actually understands anything. Classic story about technologies devouring themselves.