There was a paper in 2016 exploring how an ML model was differentiating between wolves and dogs with a really high accuracy, they found that for whatever reason the model seemed to *really* like looking at snow in images, as in thats what it pays attention to most.
Then it hit them. *oh.*
*all the images of wolves in our dataset has snow in the background*
*this little shit figured it was easier to just learn how to detect snow than to actually learn the difference between huskies and wolves. because snow = wolf*
Shit like this happens *so often*. People think trainning models is like this exact coding programmer hackerman thing when its more like, coralling a bunch of sentient crabs that can do calculus but like at the end of the day theyre still fucking crabs.











