Artificial intelligence seems to be much worse at recognising objects from poorer countries. This could have a negative impact on a variety of systems that can rely on AI, including driverless cars.
Terrance DeVries and colleagues at Facebook AI Research took six standard image recognition systems and showed them photos of 135 different types of objects. The AIs were less effective at recognising household items from non-Western and low-income countries.
In one example, an image of soap taken in the UK, where the average monthly income is $1890, was compared to one taken in Nepal, where the average monthly income is $288.
Advertisement
For the Nepal photo, which showed a bar of soap on a dish, none of the AIs recognised it as soap. Instead, suggestions included 鈥渇ood鈥�, 鈥渃heese鈥� and 鈥渟pam鈥�. For the UK photo, which was a plastic bottle of hand soap, every AI suggested something related to soap, toilet or sink, and none guessed it was food.
Image recognition systems are typically trained with millions of images, but these data sets are not very geophraphically diverse. 鈥淲hen we only show [AIs] a small slice of the world, they鈥檙e going to be unprepared for everything else,鈥� says Janelle Shane, an AI researcher.
Biases in AI are not new. However, 鈥渢his is the first time I鈥檝e seen it demonstrated in object recognition, which is incredibly important given the uses object recognition has鈥�, says Os Keyes, from the University of Washington.
These applications include systems that help people who are blind identify objects around them, as well as vision systems for self-driving cars.
鈥淎rtificial intelligence is dominated by white, male, western, middle-class perceptions of the world鈥�, says Mike Cook at Queen Mary University, London. 鈥淭his is a good, clear example of what effect that has.鈥�
搁别蹿别谤别苍肠别:听补谤齿颈惫,听
