Saturday, 27 June 2020


We’ve seen the built-in bias on display of this neural network application that turned a pixelated image of Barack Obama into an avatar that presents as pretty Caucasian, and Janelle Shane (previously) does a really good job at unpacking what’s going on here with our own tendency for pareidolia codified and amplified.
Not only is the algorithm informed by representation (and under representation) which is highly problematic and is something that the industry desperately needs to redress lest machine learning become the next commercialised embodiment of unreliability, the artificial intelligence delivers what it’s rewarded for delivering, be that a human face or a serviceable suspect that complies well enough with a blurry or grainy image. Thankfully most of the leaders in this sector, faults and all, are taking a pause in sharing their technologies with bad actors, including law enforcement agencies. The application cannot recover details that do not exist—only invent them based on what’s been judged plausible.