The Trans Dataset Built Without Permission and Stored Improperly
Facial recognition technology is here. Whether we like it or not, cameras all across the world are scanning faces and building databases.
There’s a popular misconception that technology is objective and unbiased. But that’s not true. All systems carry the biases of the people who created them, and nowhere is that more evident than in facial recognition systems.
Today’s show is about how those biases come to bear, and the dangers of running recklessly forward without considering the consequences. All the way back in 2013, the University of North Carolina, Wilmington published a dataset meant for facial recognition systems. It contained more than 1 million images of trans people, pulled from YouTube, showing them at various stages of their transition.
This was done without the permission of the original posters. Why? Because terrorists might take hormones to alter their face and beat border control systems.
It gets weirder from there.
Here to help us tell the story is Os Keyes. Keyes is a researcher and PhD candidate at the University of Washington’s Department of Human Centered Design & Engineering. They’re also the co-author of Feeling fixes: Mess and emotion in algorithmic audits, which is a scientific audit of the dataset we’re going to be talking about today.
Stories discussed in this episode:
We’re recording CYBER live on Twitch and YouTube. Watch live during the week. Follow us there to get alerts when we go live. We take questions from the audience and yours might just end up on the show.
Subscribe to CYBER on Apple Podcasts or wherever you listen to your podcasts.