Remember thinking it was just about the model until @fionajenkins pointed this out. Used to dump messy data in and wonder why things went weird. Totally changed how I start any project now.
Yeah I was just reading about how garbage data trains biased models, like that Wired piece on facial recognition failing on darker skin tones. They traced it back to totally uncurated training sets. It's wild how the whole "garbage in, garbage out" thing gets ignored until it blows up publicly.