Why AI Needs So Many Examples
Michael Running Wolf is a citizen of the Northern Cheyenne Tribe who grew up in Birney, Montana, a remote village where running water and electricity were not always reliable. He earned bachelor's and master's degrees in computer science at Montana State University, then worked as a software engineer on Amazon's Alexa voice assistant. He later taught at Northeastern University and went on to pursue a PhD in computer science at McGill University in Canada.
Voice assistants rely on automatic speech recognition — AI that turns spoken sound into words. These systems are trained on training data: enormous collections of recorded speech, often with written transcripts. The more varied examples a model hears — different voices, accents, speeds and background noise — the better it becomes at understanding someone it has never heard before.
That is exactly the problem for endangered Native languages. Michael has pointed out that for many Montana tribes, the recorded audio that exists adds up to perhaps around 100 hours, and some tribes have almost none. Mainstream speech systems learn from vastly more. Many Indigenous languages are also polysynthetic, meaning a single long word can carry the meaning of a whole English sentence. Handling that kind of word-building is still an unsolved challenge for AI.
To tackle it, Michael co-founded First Languages AI Reality (FLAIR), which develops AI tools that communities can adapt to revitalize their languages. He argues that communities need their own engineers and computer scientists, and control over their own data, so the technology serves the people whose language it is.
