State-of-the-art image-classifying AI models trained on ImageNet, a popular (but problematic) dataset containing photos scraped from the internet, automatically learn humanlike biases about race, gender, weight, and more according to new research from scientists at Carnegie Mellon University and George Washington University.