Overheard a Google engineer talking about AI bias at a coffee shop in Seattle

I was sitting near this guy at Starbucks downtown, and he was telling his friend how their training data for a new model had too much of a certain demographic, like 80% white male voices. It made me realize how easy it is to bake in biases without even knowing it. Has anyone else run into issues with their own datasets leaning one way?

2 comments

2 Comments

xenarobinson12d ago

actually i'm gonna push back on this a little... bias isn't always a bad thing. if you're building a voice model for a specific use case, like say a medical diagnosis tool for older men, then having mostly male voices in your training data makes sense. you don't need perfect representation if the end users are a narrow group. google's mistake might be trying to make everything for everyone when some stuff works better when it's focused.

clark.iris12d ago

idk, focusing too much can blind you to who else might need that tool.