A photo app's face-detection feature works well for some skin tones but poorly for others, because the training photos mostly showed one group of people. Which best describes this problem?