SoylentNews

This month, Facebook engineers and data scientists revealed the outcome of a group of experiments called “P(Bad for the World).” Posts were designated as good or bad for the world through user surveys, and the employees found that posts seen by large numbers of people were more likely to be in the bad column.

Shocked into action, the team made an AI model that could predict these negative posts and apply lower rankings to them in news feeds. Initial tests showed that the algorithm successfully limited harmful posts’ visibility, but it also decreased how often users open Facebook — a metric of great value to executives, likely tied to how often the company can serve users ads. The algorithm was ultimately changed to apply lighter demotions to a larger swath of harmful posts.

Several other proposals never made it onto the platform at all. One feature would retroactively alert users that they shared false information and another algorithm deprioritized “hate baiting” posts which seem relatively innocuous but inspire vitriolic comments.

While employees felt these proposals were knocked down because they would disproportionately affect right-wing users and publishers as well as the bottom line, Facebook’s integrity executive Guy Rosen told the NYT that these assertions were incorrect. Rosen claims the first tool simply wasn’t effective, and the hate-baiting tool could be weaponized against pages if users’ spammed the comments of their posts.

Original Submission