I only believe in statistics that I doctored myself.
There is a rampant tendency in any industry where someone is trying to sell something with a bunch of data, where they cherry pick a little bit... bias a little bit. This becomes quite easy when there is an enormous amount of data to cherry pick from.
Algorithms learn by being fed certain images, often chosen by engineers, and the system builds a model of the world based on those images. If a system is trained on photos of people who are overwhelmingly white, it will have a harder time recognizing nonwhite faces.
One of the most untruthful things possible, you know, is a collection of facts, because they can be made to appear so many different ways.
People think 'big data' avoids the problem of discrimination because you are dealing with big data sets, but, in fact, big data is being used for more and more precise forms of discrimination - a form of data redlining.
Like dreams, statistics are a form of wish fulfillment.
One thing that the Internet has created is the sense that information is at your fingertips, when it's really only a very, very limited, specific, and slanted kind of information.
We should always be suspicious when machine-learning systems are described as free from bias if it's been trained on human-generated data. Our biases are built into that training data.
