FreeFeed: Combating Unethical and Manipulative AI With Inferences From Human Interactions On Social Media

Current algorithms that are utilized on socialmedia feeds follow a basic profit-based model,which suggests posts based on previous userinteractions. However, when users areconstantly exposed to the same content, they areforced into believing the only perspective thatthey see, which makes these algorithms highlymanipulative. This can have negativeimplications on individuals suffering fromeating disorders, depressive symptoms, anddrug addiction, as they are continuouslyexposed to negative posts (ex. body-shaming,self-harm content, etc.). This study proposes arevised machine learning algorithm (FreeFeed)that will vary the feed so that it can introducemultilateral perspectives, thus allowing users tofreely formulate their own opinions afteranalyzing multiple viewpoints. At first, theTwitter API was filtered off of 4 stress-relatedrisk factors in adolescents: drugs, relationships,academics, and body image/physicalappearance. After 120,000 tweets were collected and preprocessed, the tweets wereused to train/test a generalized logisticregression model and a multi-layer perceptronneural network. The models were compared onvalues such as the F1 score (max 0.963), AUC(0.997), and accuracy (max 93.7%). Thealgorithm was then implemented into a site andtested on a set of 100 social media users in FairLawn, New Jersey, to identify FreeFeed’simpact on self-esteem. Over the course of aweek, participants completed a survey beforeand after use, in which responses were scoredon the Rosenberg Self-esteem Scale. The testsubjects were split into 3 cohorts: a control thatused pre-existing feed algorithms, a group thatutilized FreeFeed for 15 minutes per day, and agroup that used FreeFeed 30 minutes per day.After a full week of usage, individuals that usedthe FreeFeed algorithm for 30 minutes/day hada 20.46% increase in self-esteem. Overall,FreeFeed has the ability to protect billions ofindividuals from the side-effects of highlymanipulative algorithms.