Public discourse can show which consequences people associate with Artificial Intelligence (AI), which groups they expect to be affected, and whether they interpret those consequences as benefits or threats. We analyze public perceptions of AI’s social impacts across n=98,413 YouTube comments collected from 2022 to 2025 using a two-stage LLM pipeline that filters neutral and off-topic content before annotating target sentiment, fear, sarcasm, and emotion. Target sentiment achieves Quadratic Weighted Kappa (QWK) =0.738 against a human-labeled sample. In exploratory keyword-defined subsets, safety and human-control comments exhibit the highest fear (0.659) and most negative sentiment (-0.521), while comments mentioning children, families, and workers are more threat-oriented than those mentioning students and teachers. Negative sentiment remains common among the retained evaluative comments, but whether it rises or falls over time depends on how videos are weighted. Interestingly, sentiment changes near product releases are comparable to those at placebo dates. These results can help direct future research and policy attention toward the concerns most visible in public discourse, especially as a starting point for validation with affected groups.