Researchers say AI agents can develop new languages—and other agents can learn them

The researchers say they tested open- and closed-weight language models across many runs in a controlled GlossoGen scenario. They describe the emerging languages as compositional and morphologically productive: meaningful parts could be combined, and new word forms created. The team reports that new agents learned these languages by observing their use without seeing how they were constructed—including weaker models unable to develop languages on their own. Learners also repaired failed conversations through targeted questions, the researchers say.