,
8 tweets,
3 min read
Read on Twitter
There seems to be some confusion about whether algorithms—which are created by people—can have racial or gender bias. The answer is yes, yes they can. Let’s break it down. 1/
Machine learning algorithms pick up patterns in data, including when those patterns reflect societal biases. Researchers found two photo collections used in AI research had a gender bias in who was shown doing activities such as cooking and sports. 2/ wired.trib.al/7pRNg90
In 2015, a black software developer embarrassed Google by tweeting that the company’s Photos service had labeled photos of him with a black friend as “gorillas.” Google declared itself “appalled and genuinely sorry.” 3/
Part of Google’s solution was to make Photos blind to gorillas and chimps altogether, to prevent it applying those labels to people. The fact even Google can’t trust its AI algorithms shows the trickiness of this problem: algorithms lack common sense. 4/ wired.trib.al/1GRkXa4
Here’s another example of biased algorithms: Last year, researchers showed that Microsoft and IBM were selling face analysis services that were much less accurate for dark skinned women than light skinned men. 5/ wired.trib.al/msDfHvx
When the algorithms were put to the test, Microsoft’s could identify the lightest male faces in a collection of photos as men every time. When shown darker female faces, the service mis-gendered them 21 percent of the time. Both companies say they have improved their services. 6/
Why does this keep happening? One reason is that data used to train algorithms tends to reflect historical biases in society. It might also be a reflection of how the world’s pool of AI researchers is not very diverse: only an estimated 12% women. 7/ wired.trib.al/a0QngwD
Fortunately, there are many people inside & outside of the AI community working to draw attention to biased algorithms and find ways to detect and fix them. So while algorithms can absolutely be biased—just like with people we can do something about it. 8/ wired.trib.al/txcbIVC
