Feelings about Artificial Intelligence: An Emotion Detection Approach Using BERT
AbstractArtificial Intelligence (AI) is increasingly integrated into everyday life, raising concerns and expectations about its future. Traditional surveys and interviews have been conducted to understand societal feelings about AI, but they have been limited in capturing spontaneous feelings. This study utilizes social media as a data source to analyze emotions regarding AI's trajectory to obtain spontaneous and freely shared contemporary content by various segments of society. A fine-tuned BERT model, trained on the GoEmotions dataset with 28 emotion categories, was used for multiclass emotion classification. Analysis of Reddit posts and comments across AI-related subreddits spanning from 2012 to 2022 revealed a spectrum of emotions, with neutrality being the most prevalent, followed by curiosity and approval. The least common categories were relief and pride. Moreover, among these 28 emotion categories, seven categories (i.e., excitement, fear, approval, disapproval, optimism, confusion, and curiosity) were chosen and examined how their frequencies have changed over time. Spikes in fear and confusion correlated with AI advancements, such as the 2015 autonomous weapons debate and the 2017 AlphaGo victory. More recent discussions exhibited increased approval and excitement, particularly during the COVID-19 pandemic when AI applications gained prominence. The key idea of this study was to understand societal feelings and their trends from large amounts of text data and to build classifiers that can detect more than Ekman’s five emotion classes to explore richer empirical results. These insights contribute to AI policy discussions, human-centered AI innovation, and the methodological integration of computational techniques in social science research.