Can AI's Voracious Appetite Be Tamed?

In the spring of 2019, artificial intelligence datasets started disappearing from the internet. Such collections -- typically gigabytes of images, video, audio, or text data -- are the foundation for the increasingly ubiquitous and profitable form of AI known as machine learning, which can mimic various kinds of human judgments such as facial recognition. In April, it was Microsoft's MS-Celeb-1M, consisting of 10 million images of 100,000 people's faces -- many of them celebrities, as the name suggests, but also many who were not public figures -- harvested from internet sites. In June, Duke University researchers withdrew their multi-target, multi-camera dataset (DukeMTMC), which consisted of images taken from videos, mostly of students, recorded at a busy campus intersection over 14 hours on a day in 2014. Around the same time, people reported that they could no longer access Diversity in Faces, a dataset of more than a million facial images collected from the internet, released at the beginning of 2019 by a team of IBM researchers. All together, about a dozen AI datasets vanished -- hastily scrubbed by their creators after researchers, activists, and journalists exposed an array of problems with the data and the ways it was used, from privacy, to race and gender bias, to issues with human rights.