This study explores public attitudes towards artificial intelligence (AI) by analyzing user discussions on Telegram chats. The analysis integrates Latent Dirichlet Allocation (LDA) for topic discovery and an emotion classificator that was initially trained on a labeled subset of the data. Thus, the sentiment analysis process involved a two-stage method, starting with preliminary data annotation using GPT-4o followed by emotion classification through supervised learning. The paper reports on experimental comparisons of different vectorization techniques and classification models to identify the most effective approach for emotion recognition within AI-related conversations. The findings uncover five primary themes in the discourse about AI, including general opinions, accessibility issues, creative uses, and technical performance evaluations. Sentiment analysis indicates a prevalent sense of surprise and curiosity toward AI technologies, coupled with persistent undertones of sadness and fear. The approach demonstrates the utility of combining language models with interpretable classifiers and highlights the potential of discourse-based methods for understanding affective engagement with emerging technologies.