Scientists decided to test AI's ability to identify patient characteristics from X-rays, Popmech reported.
Before the experiment, the researchers trained the AI on three large sets of chest X-ray data that had captions identifying the person's race - dark-skinned, white-skinned, or Asian.
The scientists then showed X-rays of many different body regions of different groups of people and asked the AI to determine the person's race. The pictures lacked any indication of race: no skin color, texture, or hair color.
For all body parts, the AI was able to determine race with about 90% accuracy, even without any identifying features.
The scientists decided that the AI somehow used statistical data and based its conclusions on factors such as body mass index or density/chest circumference. They removed possible associated variables and showed only data sets of people with similar BMI and body types, but still the AI was able to determine race.
The researchers don't know how it does it. One assumption is that the AI recognizes differences in melanin in people's skin that show up in pictures and are not visible to the eye.