It seems even artificial intelligence (AI) falls victim to the oldest flaw in computers since history, GIGO; Garbage In, Garbage Out. Even the AI, will be as smart as you program it to be, or as dumb as it was programmed.
We all imagine what the future will be like with self-driving cars. Looking at how Uber and the likes have disrupted commuting, just think of how much more efficient and affordable rides will be when you put a computer behind the wheels.
Well, researchers at the Georgia Institute of Technology imagined just that; and went as far as using eight AI models that are to be used in the state-of-the-art object detection system.
That is systems that would allow self-driving cars to detect road signs, pedestrians, among other things on the roads and parking lots. It turns out, the cars were significantly poor in detecting people with very dark skin, and the lighter the pedestrian was, the better it was at detecting them.
The researchers conducted their experiments using pedestrians of different shades of skin color. The pedestrians were classified according to the Fitzpatrick scale, which is commonly used as a human skin color classification board.
The researchers found out that the AI behind the wheel showed “uniformly poorer performance” when the pedestrians hade the three darkest shades on the spectrum. On average, the AI accuracy decreased by 5% when the skin tone got darker.
That is even after the researchers had accounted for variables such as time of the day and night; with reference to lighting. The researchers concluded that before AI-powered self-driving cars become a thing, the darker skinned pedestrians might more often fall victim to these cars on the roads compared to their lighter-skinned counterparts.
To avoid such a scenario, the researchers say the self-driving cars must first be given lots of images of darker skinned pedestrians in the data sets to improve their recognition of the various shades of skin color.