Analysis of conspiracy-based content in social engineering with artificial intelligence

With the rapid advancement of technology, artificial intelligence (AI) has increasingly become a powerful force in modern life, influencing the ways individuals interact with digital platforms and consume information. Although these technological developments have made daily tasks easier and more efficient, they have also introduced new risks and challenges, particularly in the realm of social media. The rise of the digital age and widespread use of online platforms have transformed social media into not only a communication tool but also a space where misinformation, emotional engagement, and conspiracy theories spread. Therefore, analyzing emotional responses in these environments and understanding user behavior have become important areas of academic research. This thesis aims to explore how artificial intelligence can be used to detect and analyze users' emotional states on social media platforms, especially within the context of conspiracy theories. The primary focus of the study is sentiment analysis using advanced machine learning models to determine whether social media content carries positive, negative, or neutral sentiments. Additionally, the psychological effects of exposure to misleading information are examined, along with how such emotional content shapes public discourse. To this end, a publicly available dataset from Kaggle containing text data collected from platforms like Twitter, Facebook, and Instagram was utilized. These platforms were selected due to their high user interaction and the prevalence of emotionally charged and controversial content. The study conducted a detailed analysis of user interactions, focusing particularly on the emotional tone and responses associated with conspiracy theories and frequently circulated misinformation in digital spaces. Before model training, comprehensive data preprocessing is performed to ensure the text data conforms to its standards and features. Before processing, punctuation processing ensures that all information is converted to lowercase, punctuation units are removed, and stemming and rooting techniques can be run. These natural language processing (NLP) methods clean and structure the raw data, making it suitable for machine learning. This process not only reduces noise and redundancy in the dataset but also improves the model's generalization ability by focusing on key meaningful details in the text. As a result, the data is more standardized, allowing for more efficient annotation and feature extraction in subsequent steps. Furthermore, this type of preprocessing minimizes the risk of data sparsity and increases the overall range of training by reducing programming complexity. Three different machine learning models were employed for sentiment analysis: Naive Bayes, LSTM (Long Short-Term Memory), and BERT (Bidirectional Encoder Representations from Transformers). Each model was trained and tested within the same experimental framework, and their performances were objectively compared. Naive Bayes was chosen as a baseline model due to its simplicity and speed but showed limitations with complex sentence structures. LSTM, designed to handle sequential data, performed better than Naive Bayes, especially on longer texts. However, BERT significantly outperformed both models, achieving the highest accuracy, precision, recall, and F1 score. The results demonstrated that BERT was particularly effective at detecting subtle emotional expressions in social media texts, thanks to its ability to understand context and word relationships bidirectionally. This provided a major advantage for sentiment detection in conspiracy content, where language is often indirect, sarcastic, or emotionally manipulative. The study confirmed that BERT is the most reliable model for identifying nuanced emotional tones linked to sensitive content and psychological reactions. The performance of each model was evaluated using common classification metrics such as Precision, Recall, and F1-Score. These metrics allowed a detailed comparison of the models' effectiveness in recognizing and categorizing sentiment in social media texts. The success of the BERT model highlighted the importance of context-aware, transformer-based architectures in complex sentiment analysis tasks. Beyond technical analyses, the thesis also examined the psychological implications of the findings. It was observed that social media platforms tend to amplify negative emotions in reactions to content related to conspiracy theories. Users exposed to such narratives frequently express feelings of fear, anger, distrust, and confusion. If left unaddressed, these emotional responses can contribute to broader societal issues such as polarization, anxiety, and acceptance of misinformation. Based on these findings, the thesis offers several key recommendations. First, it emphasizes increasing digital media literacy. Improving users' ability to recognize emotional manipulation and misinformation can help mitigate the negative effects of exposure to harmful content. Educational programs that encourage critical thinking when interacting with online media can encourage more responsible engagement. Providing digital literacy education, especially for young students, at an early age can be found to foster "more informed and resilient growth." Second, the study advocates for stronger algorithmic accountability on social media platforms. Content recommendation systems that prioritize popularity over accuracy or emotional well-being often promote sensational and emotionally charged content, disrupting the balance of the information ecosystem. Therefore, it is essential for platforms to develop more ethical algorithms that evaluate not only user engagement but also the accuracy and emotional impact of content. Emotion-based filtering systems and content labeling tools can help users experience healthier digital interactions. Third, the study recommends the development of AI-powered educational tools aimed at increasing users' awareness of their online behavior. These tools could encourage more conscious and responsible participation in digital environments by providing real-time feedback on the emotional tone of user-generated content or interactions. Such an approach could help users become more aware of their own emotional reactions while also fostering greater empathy for others in online environments. The research also recommends incorporating multimedia formats such as video, images, and audio into future studies, given their significant role in emotional communication on social media. Moving beyond text-based content and integrating visual and audio data into sentiment analysis could lead to more comprehensive insights. In this context, the development of multimodal analysis techniques could enable more accurate and holistic assessments of sentiment across platforms. Finally, the thesis encourages interdisciplinary collaboration between AI researchers and mental health professionals to investigate how emotional patterns emerging on social media can be used in early intervention and support strategies. Collaborative efforts in this area could lead to the development of innovative solutions that support mental health in digital environments. Early detection and effective addressing of signs of online trauma, stress, or loneliness can be crucial for protecting users' psychological well-being. In conclusion, this research demonstrates the powerful potential of advanced natural language processing models (specifically BERT) for accurately performing sentiment analysis in social media environments. The study further explores the impact of digital evolution and its psychological implications by identifying emotional patterns in online narratives that include conspiracy theories. By combining technical program, social analysis, and practical recommendations, it makes a lasting contribution not only to the field of Natural Language Processing (NLP) but also to digital communication ethics and mental health. Ultimately, a more informed, emotionally intelligent, and resilient digital society is essential.