Identifying Online Conspiracy Theories: A Multiphase Analysis of AI Models

The growth of online platforms has amplified the spread of conspiracy theories, presenting a significant challenge to social cohesion and online safety. While platforms like Twitter and Facebook have received considerable academic attention, the unique ecosystem of YouTube comments remains comparatively underexplored as a site of conspiratorial discourse. This dissertation addresses this gap by presenting a systematic investigation into the detection of conspiratorial narratives within YouTube comments using a comparative evaluation of artificial intelligence techniques. This dissertation presents and validates a multiphase methodological framework through a systematic empirical investigation. The research first conducts an exploratory discourse analysis of a prominent case study, the 2023 Maui wildfires, to identify the distinct linguistic, emotional and topical signatures of conspiratorial content. Building on these findings, a novel dataset of over 7,300 YouTube comments is manually annotated. The core of the study is a rigorous comparative evaluation of three tiers of AI detection methods: (1) traditional machine learning models (e.g., SVM, XGBoost) using engineered text and psycholinguistic features; (2) fine-tuned Transformer models (BERT, BERTweet); and (3) zero-shot classification with Large Language Models (LLMs). The results reveal a clear hierarchy of performance, with LLMs demonstrating superior accuracy, achieving a top F1-score of 0.880. Notably, a hybrid approach combining Google’s foundation model embeddings with an XGBoost classifier also performed exceptionally well, outperforming the fine-tuned Transformer models. A key finding is that the explicit addition of sentiment and emotion features did not consistently improve model performance. This suggests that the rich, high-dimensional embeddings from modern foundation models already implicitly encode these psycholinguistic signals, rendering separate feature engineering mostly redundant. This research concludes that the effective detection of nuanced, narrative-driven phenomena like conspiracy theories is increasingly dependent on deep semantic understanding rather than surface-level linguistic markers. The findings hold significant implications for both computational social science and practical content moderation,advocating for a shift towards leveraging foundation models and suggesting a prag�matic, structured approach for developing scalable moderation systems. Keywords: Conspiracy Theories, Disinformation, YouTube, Natural Language Processing (NLP), Large Language Models (LLMs), Machine Learning, Computational Social Science