This article describes the efforts of Facebook, Youtube, and similar hosts of user-generated content, to screen unacceptable material. (Both speech and images.) It’s apparently a grim task, because of the depravity of some material. For the first decade, moderation methods were heavily ad hoc, but gradually grew more complex and formalized in response to questions such as when to allow violent images as news. In aviation terms, it was at Stage 2: Rules + Instruments. Now, some companies are developing Stage 3 (standard procedures) and Stage 4 (automated) methods.
In May 2014, Dave Willner and Dan Kelmenson, a software engineer at Facebook, patented a 3D-modeling technology for content moderation designed around a system that resembles an industrial assembly line. “It’s tearing the problem [of huge volume] into pieces to make chunks more comprehensible,” Willner says. First, the model identifies a set of malicious groups — say neo-Nazis, child pornographers, or rape promoters. The model then identifies users who associate with those groups through their online interactions. … This way, companies can identify additional or emerging malicious online activity. “If the moderation system is a factory, this approach moves what is essentially piecework toward assembly,” he said. “And you can measure how good the system is.”
MODERATION IS “A PROFOUNDLY HUMAN DECISION-MAKING PROCESS.”