Can AI Rescue Us From Violent Images Online?

A missile fired by Iraqi government forces. Last week I wrote about how deep learning image recognition algorithms offer a potential solution to the epidemic of violent imagery cascading across the online world, especially with respect to the rising issue of live streaming violent acts that can reach vast audiences before any human reviewer has a chance to examine the footage. What might this actually look like in practice? For the past year my open data GDELT Project has been applying Google's Cloud Vision API (its cloud-based deep learning image cataloging service) to global news imagery from every country in the world, totaling more than a quarter billion images to date. While no algorithm is perfect and the Vision API does make mistakes, it has proven remarkably adept at recognizing an incredible wealth of violence-related indicators from imagery spanning the entire globe.