Why would the Internet lie to me?: Analyzing the Performance of Misinformation on Twitter utilizing Large Language Models, Machine Learning, and Evolutionary Computing

Abstract

Misinformation on social media continues to be an issue for social media users with an increasing number of problems and threats seemingly stemming from the spread of deceptive and usually politically motivated information. This misinformation or fake news has been consistently shown to have the potential to damage and change institutions, democracy, and people’s lives on social media. While there has been research into why specific pieces of disinformation spread, we are interested in determining if we can artificially replicate and improve the performance of tweets containing misinformation. Through this, we can then analyze why specific tweets or specific pieces of misinformation spread farther than others. We achieve this through developing an evolutionary algorithm that automatically optimizes the virality of tweets. We use a deep learning model to predict the number of retweets a specific text may receive. This serves as our measure of fitness in the evolutionary algorithm, as it is our closest analog to how far a tweet may spread. Using this prediction, we can then leverage Large Language Models (LLMs) to mutate and crossover disinformation tweets in an evolutionary iteration. Based on the fitness of these mutated and changed tweets, we can then see differences in performance and analyze why that may be. Through this, we can then pinpoint reasons as to why specific pieces of misinformation spread and utilize this knowledge to help reduce the spread of such misinformation.